chore: adding the data and figure procssing

Baseline comparisons

.dockerignore

@@ -0,0 +1,17 @@
+.git
+.venv
+.venv-tpu
+**/__pycache__
+**/*.pyc
+**/*.pyo
+**/.pytest_cache
+**/.mypy_cache
+**/.ruff_cache
+**/.ipynb_checkpoints
+wandb
+build
+paper/build
+paper/build-cais
+node_modules
+**/node_modules
+*.egg-info

.env.sweep.example

@@ -0,0 +1,24 @@
+# Copy this file to .env.sweep and fill in values.
+
+# Required for wandb runs and sweep agent workers.
+WANDB_API_KEY=
+WANDB_ENTITY=
+WANDB_PROJECT=capstone
+
+# Required for private repo bootstrap workers.
+GITHUB_TOKEN=
+
+# Optional defaults for bootstrap mode.
+# REPO_URL=https://github.com/org/repo.git
+# BRANCH=main
+# WORKDIR=$HOME/PHANTOM-agent
+# SWEEP_ID=entity/project/id
+# AGENT_COUNT=0
+# AGENT_LOOP=1
+# RETRY_SECONDS=20
+
+# Optional local benchmark defaults.
+# LOCAL_BENCHMARK_ARGS=--tiers static,surge,linear,qtable,ppo --alpha-values 0.0,0.3 --episodes 3 --total-timesteps 3000 --max-steps 40 --device cpu
+# SIMPLE_BENCHMARK_ARGS=--tiers qtable,ppo,dqn,a2c --alpha-values 0.0,0.15,0.3,0.45,0.6 --episodes 8 --total-timesteps 8000 --max-steps 40 --device cpu
+# PHANTOM_BENCHMARK_COMPARE_ROBUST=1
+# BENCHMARK_AGENT_ARGS=--tiers static,surge,linear,qtable,ppo --alpha-values 0.0,0.3,0.6 --episodes 5

.github/workflows/latex.yml

@@ -12,17 +12,168 @@ on:
 jobs:
   build:
     runs-on: ubuntu-latest
+    env:
+      OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+      R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+      R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+      R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+      R2_BUCKET_NAME: ${{ secrets.R2_BUCKET_NAME }}
     steps:
       - uses: actions/checkout@v4
-      - name: Compile LaTeX document
+
+      - name: Prepare appendix code snapshot
+        run: bash paper/concat_code.sh
+
+      - name: Generate mirrors with Codex
+        if: ${{ env.OPENAI_API_KEY != '' }}
+        uses: openai/codex-action@v1
+        with:
+          openai-api-key: ${{ env.OPENAI_API_KEY }}
+          sandbox: workspace-write
+          safety-strategy: drop-sudo
+          working-directory: .
+          prompt: |
+            Read and follow the mirror instructions in `paper/src/mirrors/genpop/INSTRUCTIONS.md`.
+
+            Source chapters are in `paper/src/chapters/`:
+            - 01-intro.tex
+            - 02-literature-review.tex
+            - 03-methodology.tex
+            - 04-results.tex
+            - 05-discussion.tex
+            - 06-conclusion.tex
+
+            Update `paper/src/mirrors/genpop/*.tex` so they mirror the thesis for a general audience according to the instruction file.
+            Keep LaTeX valid and preserve citation commands and section order.
+
+            Then create or update `paper/src/main-mirror-genpop.tex` by using `paper/src/main.tex` as the base and replacing chapter inputs from `chapters/...` to `mirrors/genpop/...`.
+            Do not change any other project files.
+
+      - name: Compute LaTeX roots
+        id: roots
+        run: |
+          {
+            echo "root_files<<EOF"
+            echo "main.tex"
+            for file in paper/src/main-mirror-*.tex; do
+              if [ -f "$file" ]; then
+                basename "$file"
+              fi
+            done
+            echo "EOF"
+          } >> "$GITHUB_OUTPUT"
+
+          echo "Compiling roots:"
+          echo "main.tex"
+          for file in paper/src/main-mirror-*.tex; do
+            if [ -f "$file" ]; then
+              basename "$file"
+            fi
+          done
+
+      - name: Compile LaTeX documents
         uses: xu-cheng/latex-action@v3
         with:
-          root_file: main.tex
+          root_file: ${{ steps.roots.outputs.root_files }}
           working_directory: paper/src
-          args: -pdf -interaction=nonstopmode -file-line-error -outdir=../build
-          pre_compile: bash ../concat_code.sh
-      - name: Upload PDF
+          args: -pdf -f -interaction=nonstopmode -file-line-error -r ../.latexmkrc -outdir=../build
+
+      - name: Upload PDF artifacts
         uses: actions/upload-artifact@v4
         with:
           name: thesis-pdf
-          path: paper/build/main.pdf
+          path: |
+            paper/build/main.pdf
+            paper/build/main-mirror-*.pdf
+
+      - name: Get current date
+        id: date
+        run: echo "date=$(date +'%Y-%m-%d')" >> $GITHUB_OUTPUT
+
+      - name: Upload to Cloudflare R2
+        if: ${{ env.R2_ACCESS_KEY_ID != '' && env.R2_SECRET_ACCESS_KEY != '' && env.R2_ENDPOINT != '' && env.R2_BUCKET_NAME != '' }}
+        env:
+          AWS_ACCESS_KEY_ID: ${{ env.R2_ACCESS_KEY_ID }}
+          AWS_SECRET_ACCESS_KEY: ${{ env.R2_SECRET_ACCESS_KEY }}
+          AWS_ENDPOINT_URL: ${{ env.R2_ENDPOINT }}
+          DATE: ${{ steps.date.outputs.date }}
+          BUCKET_NAME: ${{ env.R2_BUCKET_NAME }}
+        run: |
+          pip install boto3
+          python3 << 'EOF'
+          import boto3
+          import os
+
+          s3 = boto3.client('s3',
+              endpoint_url=os.environ['AWS_ENDPOINT_URL'],
+              aws_access_key_id=os.environ['AWS_ACCESS_KEY_ID'],
+              aws_secret_access_key=os.environ['AWS_SECRET_ACCESS_KEY']
+          )
+
+          date = os.environ['DATE']
+          bucket = os.environ['BUCKET_NAME']
+
+          # upload dated version
+          dated_filename = f"thesis-{date}.pdf"
+          s3.upload_file(
+              'paper/build/main.pdf',
+              bucket,
+              dated_filename,
+              ExtraArgs={'ContentType': 'application/pdf'}
+          )
+          print(f"Uploaded {dated_filename}")
+
+          # upload latest version
+          s3.upload_file(
+              'paper/build/main.pdf',
+              bucket,
+              'thesis-latest.pdf',
+              ExtraArgs={'ContentType': 'application/pdf'}
+          )
+          print(f"Uploaded thesis-latest.pdf")
+
+          # upload mirror versions (if generated)
+          build_dir = 'paper/build'
+          for filename in os.listdir(build_dir):
+              if not filename.startswith('main-mirror-') or not filename.endswith('.pdf'):
+                  continue
+              mirror_name = filename[len('main-mirror-'):-4]
+              source_path = os.path.join(build_dir, filename)
+
+              dated_mirror = f"thesis-{mirror_name}-{date}.pdf"
+              latest_mirror = f"thesis-{mirror_name}-latest.pdf"
+              namespaced_dated = f"mirrors/{mirror_name}/thesis-{date}.pdf"
+              namespaced_latest = f"mirrors/{mirror_name}/thesis-latest.pdf"
+
+              s3.upload_file(
+                  source_path,
+                  bucket,
+                  dated_mirror,
+                  ExtraArgs={'ContentType': 'application/pdf'}
+              )
+              print(f"Uploaded {dated_mirror}")
+
+              s3.upload_file(
+                  source_path,
+                  bucket,
+                  latest_mirror,
+                  ExtraArgs={'ContentType': 'application/pdf'}
+              )
+              print(f"Uploaded {latest_mirror}")
+
+              s3.upload_file(
+                  source_path,
+                  bucket,
+                  namespaced_dated,
+                  ExtraArgs={'ContentType': 'application/pdf'}
+              )
+              print(f"Uploaded {namespaced_dated}")
+
+              s3.upload_file(
+                  source_path,
+                  bucket,
+                  namespaced_latest,
+                  ExtraArgs={'ContentType': 'application/pdf'}
+              )
+              print(f"Uploaded {namespaced_latest}")
+          EOF

.gitignore

@@ -1,8 +1,92 @@
+# environment and secrets
 **/.env
+.env.*
+!.env.*.example
 **/.venv
+**/.venv-ray
+
+# python build/cache artifacts
 **/__pycache__
+phantom.egg-info/
+*.egg-info/
+
+# notebook artifacts
 **/.ipynb_checkpoints/
 **/.virtual_documents/
+
+# editor/tool state
+**/.pdf-view-restore
+.nextstep
+.ignore-gitlogue
+.cloudflare
+.nx/
+node_modules/
+dist/
+
+# generated svg/graphics
 **/session_*.svg
 **/*graph.svg
+**/auto/*.el
+
+# misc generated
+*.old
+**/package-lock.json
+**/*.parquet
+**/_build/
+
+# paper build artifacts
 paper/src/bib/auto
+paper/src/auto/*
+paper/src/bib/auto
+paper/template/*
+paper/build-cais/
+paper/defense/manim/media/
+paper/defense/manim/.manim/
+paper/src/main.pdf
+paper/src/main-blx.bib
+paper/src/svg-inkscape/
+paper/variations/
+paper/src/graphics/test_*.png
+thesis-latest.pdf
+
+# experiment run artifacts and logs
+docs/goals/*.md
+PHANTOM.wiki/
+experiments/airflow/logs/*
+experiments/airflow/logs/scheduler/
+experiments/airflow/logs/dag_processor_manager/
+experiments/collected_data/
+experiments/agents/collected_data/
+tests/e2e/test-results/
+tests/e2e/node_modules/**
+
+# rl/sim run outputs
+sim/rl/behavior_loader/*.dot
+sim/rl/behavior_loader/*.png
+sim/rl/behavior_loader/*.svg
+sim/rl/behavior_loader/*.pdf
+sim/rl/runs/
+lab/case/thesis/runs*/
+sim/case/thesis_simplified/runs*/
+
+# model binaries
+engine/models/*.zip
+engine/studies/results/*
+*.zip
+
+# wandb local state
+wandb/
+
+# data directory (large datasets)
+data/
+
+# ktem local app data
+ktem_app_data/
+
+# generated visualization pdfs
+*_mdp_viz.pdf
+phantom_env_comparison.png
+sim/phantom_env_comparison.png
+
+# web clone
+PHANTOM_web/*

.rayignore

@@ -0,0 +1,35 @@
+# Virtual environments
+.venv
+.venv*
+venv
+venv*
+**/.venv
+**/venv
+**/node_modules
+node_modules/
+
+# Python caches
+__pycache__/
+*.pyc
+.ruff_cache/
+.pytest_cache/
+
+# Git
+.git/
+
+# Large data and logs
+data/
+experiments/
+wandb/
+dumplogs*
+*.zip
+*.pdf
+*.log
+*.dot
+
+# Other large dirs
+PHANTOM_web/
+web/
+docs/
+paper/
+.nx/

AGENTS.md

@@ -0,0 +1 @@
+CLAUDE.md

Makefile

@@ -8,45 +8,236 @@ VENV      := .venv
 PYTHON    := $(VENV)/bin/python
 PIP       := $(VENV)/bin/pip
 PYTEST    := $(VENV)/bin/pytest
+NX        := npx nx
+
+SWEEP_ENV_FILE ?= .env.sweep
+TPU_CONF ?= tpu_orchestration/configs/v4_spot_us.conf
+
+WANDB_ENTITY ?=
+WANDB_PROJECT ?= capstone
+SWEEP_ID ?=
+LOCAL_TRAIN_ARGS ?= --algo ppo --total-timesteps 50000
+LOCAL_BENCHMARK_ARGS ?= --tiers static,surge,linear,qtable,ppo --alpha-values 0.0,0.3 --episodes 3 --total-timesteps 3000 --max-steps 40 --device cpu
+SIMPLE_BENCHMARK_ARGS ?= --tiers qtable,ppo,dqn,a2c --alpha-values 0.0,0.15,0.3,0.45,0.6 --episodes 8 --total-timesteps 8000 --max-steps 40 --device cpu
+BENCHMARK_AGENT_ARGS ?=
+AGENT_COUNT ?= 0
+
+WHOCLICKED_REPO ?= velocitatem/whoclickedit
+WHOCLICKED_CSV ?= experiments/exports/whoclicked.csv
+WHOCLICKED_CARD ?= experiments/exports/whoclicked_dataset_card.md
+WHOCLICKED_CSV_PATH_IN_REPO ?= whoclicked.csv
+WHOCLICKED_CARD_PATH_IN_REPO ?= README.md
+WHOCLICKED_DATASET_MESSAGE ?= Update flattened whoclickedit dataset
+WHOCLICKED_CARD_MESSAGE ?= Update dataset card for whoclickedit
+
+REPO_URL ?=
+BRANCH ?= main
+WORKDIR ?= $(HOME)/PHANTOM-agent
+AGENT_LOOP ?= 1
+RETRY_SECONDS ?= 20
+
+TRAIN_IMAGE_REF := us-central1-docker.pkg.dev/phantom-trc/phantom/phantom-trainer
+
+SWEEP_ENV_LOAD = set -a; [ -f "$(SWEEP_ENV_FILE)" ] && . "$(SWEEP_ENV_FILE)" || true; set +a
 
 .DEFAULT_GOAL := help
 
-all: pdf
-
-run.webapp:
-	@cd web && npm install && npm run dev
+.PHONY: help
+help:
+	@echo "pdf.build pdf.watch pdf.clean pdf.genpop pdf.genpop.watch pdf.arxiv | test.backend test.e2e test.all | web.dev | install | train | benchmark | benchmark.simple | benchmark.agent | train.agent | train.bootstrap | stats.lines | manim.render manim.render.all"
+	@echo "backend.server backend.provider backend.worker | platform.up platform.down platform.logs | docker.train.publish"
+	@echo "data.pull data.push data.whoclicked.publish | study.margin-erosion study.margin-erosion.quick study.margin-erosion.plot"
+	@echo "tpu.ray.bootstrap tpu.ray.deps tpu.ray.verify tpu.ray.teardown"
+	@echo ""
+	@echo "Build general public version:"
+	@echo "  make pdf.genpop"
+	@echo ""
+	@echo "Local wandb run:"
+	@echo "  make train LOCAL_TRAIN_ARGS='--algo ppo --total-timesteps 50000'"
+	@echo ""
+	@echo "Local benchmark run:"
+	@echo "  make benchmark LOCAL_BENCHMARK_ARGS='--tiers static,surge,linear --alpha-values 0.0,0.3 --episodes 3 --no-wandb'"
+	@echo ""
+	@echo "Simple benchmark run (.env.sweep defaults, robust+no_robust compare by default):"
+	@echo "  make benchmark.simple"
+	@echo ""
+	@echo "Local sweep agent from this repo:"
+	@echo "  make train.agent SWEEP_ID=entity/project/id AGENT_COUNT=5"
+	@echo ""
+	@echo "Bootstrap private repo worker from anywhere:"
+	@echo "  make train.bootstrap REPO_URL=https://github.com/org/repo.git BRANCH=main SWEEP_ID=entity/project/id"
+	@echo ""
+	@echo "Bootstrap Ray on TPU slice from config:"
+	@echo "  make tpu.ray.bootstrap TPU_CONF=tpu_orchestration/configs/v4_spot_us.conf"
+	@echo ""
+	@echo "Publish whoclickedit dataset + card:"
+	@echo "  make data.whoclicked.publish HF_TOKEN=... WHOCLICKED_REPO=velocitatem/whoclickedit"
+	@echo ""
+	@echo "Config source: $(SWEEP_ENV_FILE) (auto-loaded)"
 
 $(BUILDDIR):
 	mkdir -p paper/$(BUILDDIR)
 
-pdf: $(BUILDDIR)
-	@echo "Concatenating source code..."
-	@bash paper/concat_code.sh
-	@cd $(SRCDIR) && \
-	$(LATEXMK) -pdf -jobname=$(JOBNAME) \
-		-interaction=nonstopmode -file-line-error \
-		-outdir=../$(BUILDDIR) $(TEX)
+.PHONY: pdf.build
+pdf.build:
+	@$(NX) run paper:build
 
-watch: $(BUILDDIR)
-	@cd $(SRCDIR) && \
-	$(LATEXMK) -pvc -pdf -jobname=$(JOBNAME) \
-		-interaction=nonstopmode -file-line-error \
-		-r ../.latexmkrc \
-		-outdir=../$(BUILDDIR) $(TEX)
+.PHONY: pdf.watch
+pdf.watch:
+	@$(NX) run paper:watch
 
-clean:
-	@cd $(SRCDIR) && \
-	$(LATEXMK) -C -jobname=$(JOBNAME) -outdir=../$(BUILDDIR) || true
-	rm -rf paper/$(BUILDDIR)/*
+.PHONY: pdf.clean
+pdf.clean:
+	@$(NX) run paper:clean
+
+.PHONY: pdf.genpop
+pdf.genpop:
+	@bash scripts/nx_paper.sh build-genpop
+
+.PHONY: pdf.genpop.watch
+pdf.genpop.watch:
+	@bash scripts/nx_paper.sh watch-genpop
+
+.PHONY: pdf.arxiv
+pdf.arxiv:
+	@bash scripts/nx_paper.sh build-arxiv
+
+.PHONY: test.backend
+test.backend:
+	@$(NX) run research:test
+
+.PHONY: test.e2e
+test.e2e:
+	@$(NX) run e2e:test
+
+.PHONY: test.all
+test.all:
+	@$(NX) run-many -t test --projects=research,e2e --parallel=1
+
+.PHONY: web.dev
+web.dev:
+	@$(NX) run web:dev
 
 $(VENV):
 	python3 -m venv $(VENV)
 	$(PIP) install --upgrade pip
 
-install: $(VENV)
-	$(PIP) install -r requirements.txt
+.PHONY: install
+install:
+	@$(NX) run research:install
 
-test: $(VENV)
-	$(PYTEST) -v
+.PHONY: train
+train:
+	@WANDB_ENTITY="$(WANDB_ENTITY)" WANDB_PROJECT="$(WANDB_PROJECT)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" LOCAL_TRAIN_ARGS="$(LOCAL_TRAIN_ARGS)" $(NX) run research:train
 
-.PHONY: all pdf clean watch run.webapp install test
+.PHONY: benchmark
+benchmark:
+	@WANDB_ENTITY="$(WANDB_ENTITY)" WANDB_PROJECT="$(WANDB_PROJECT)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" LOCAL_BENCHMARK_ARGS="$(LOCAL_BENCHMARK_ARGS)" $(NX) run research:benchmark
+
+.PHONY: benchmark.simple
+benchmark.simple:
+	@WANDB_ENTITY="$(WANDB_ENTITY)" WANDB_PROJECT="$(WANDB_PROJECT)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" SIMPLE_BENCHMARK_ARGS="$(SIMPLE_BENCHMARK_ARGS)" PHANTOM_BENCHMARK_COMPARE_ROBUST="$(PHANTOM_BENCHMARK_COMPARE_ROBUST)" $(NX) run research:benchmark-simple
+
+.PHONY: benchmark.agent
+benchmark.agent:
+	@WANDB_ENTITY="$(WANDB_ENTITY)" WANDB_PROJECT="$(WANDB_PROJECT)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" SWEEP_ID="$(SWEEP_ID)" AGENT_COUNT="$(AGENT_COUNT)" BENCHMARK_AGENT_ARGS="$(BENCHMARK_AGENT_ARGS)" $(NX) run research:benchmark-agent
+
+.PHONY: train.agent
+train.agent:
+	@WANDB_ENTITY="$(WANDB_ENTITY)" WANDB_PROJECT="$(WANDB_PROJECT)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" SWEEP_ID="$(SWEEP_ID)" AGENT_COUNT="$(AGENT_COUNT)" $(NX) run research:train-agent
+
+.PHONY: train.bootstrap
+train.bootstrap:
+	@WANDB_ENTITY="$(WANDB_ENTITY)" WANDB_PROJECT="$(WANDB_PROJECT)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" REPO_URL="$(REPO_URL)" BRANCH="$(BRANCH)" WORKDIR="$(WORKDIR)" SWEEP_ID="$(SWEEP_ID)" AGENT_COUNT="$(AGENT_COUNT)" AGENT_LOOP="$(AGENT_LOOP)" RETRY_SECONDS="$(RETRY_SECONDS)" $(NX) run research:train-bootstrap
+
+.PHONY: tpu.ray.bootstrap tpu.ray.deps tpu.ray.verify tpu.ray.teardown
+tpu.ray.bootstrap:
+	@TPU_CONF="$(TPU_CONF)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" $(NX) run research:tpu-ray-bootstrap
+
+tpu.ray.deps:
+	@TPU_CONF="$(TPU_CONF)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" $(NX) run research:tpu-ray-deps
+
+tpu.ray.verify:
+	@TPU_CONF="$(TPU_CONF)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" $(NX) run research:tpu-ray-verify
+
+tpu.ray.teardown:
+	@TPU_CONF="$(TPU_CONF)" SWEEP_ENV_FILE="$(SWEEP_ENV_FILE)" $(NX) run research:tpu-ray-teardown
+
+.PHONY: data.pull data.push
+data.pull:
+	python scripts/hf_data.py pull
+
+data.push:
+	python scripts/hf_data.py push
+
+.PHONY: data.whoclicked.publish
+data.whoclicked.publish:
+	@HF_TOKEN="$(HF_TOKEN)" WHOCLICKED_REPO="$(WHOCLICKED_REPO)" WHOCLICKED_CSV="$(WHOCLICKED_CSV)" WHOCLICKED_CARD="$(WHOCLICKED_CARD)" WHOCLICKED_CSV_PATH_IN_REPO="$(WHOCLICKED_CSV_PATH_IN_REPO)" WHOCLICKED_CARD_PATH_IN_REPO="$(WHOCLICKED_CARD_PATH_IN_REPO)" WHOCLICKED_DATASET_MESSAGE="$(WHOCLICKED_DATASET_MESSAGE)" WHOCLICKED_CARD_MESSAGE="$(WHOCLICKED_CARD_MESSAGE)" $(NX) run research:whoclicked-publish
+
+.PHONY: stats.lines
+stats.lines:
+	@$(NX) run research:stats
+
+.PHONY: study.margin-erosion
+study.margin-erosion:
+	python -m engine.studies.margin_erosion_alpha
+
+.PHONY: study.margin-erosion.quick
+study.margin-erosion.quick:
+	python -m engine.studies.margin_erosion_alpha --quick
+
+.PHONY: wordcount
+wordcount:
+	@$(NX) run paper:wordcount
+
+.PHONY: docker.train.publish
+docker.train.publish:
+	@TRAIN_IMAGE_REF="$(TRAIN_IMAGE_REF)" $(NX) run research:docker-train-publish
+
+.PHONY: backend.server backend.provider backend.worker platform.up platform.down platform.logs
+backend.server:
+	@$(NX) run backend-server:dev
+
+backend.provider:
+	@$(NX) run pricing-provider:dev
+
+backend.worker:
+	@$(NX) run backend-worker:dev
+
+platform.up:
+	@$(NX) run platform:up
+
+platform.down:
+	@$(NX) run platform:down
+
+platform.logs:
+	@$(NX) run platform:logs
+
+.PHONY: pdf clean watch run.webapp test count-lines all
+pdf:
+	@$(NX) run paper:build
+
+clean:
+	@$(NX) run paper:clean
+
+watch:
+	@$(NX) run paper:watch
+
+run.webapp:
+	@$(NX) run web:dev
+
+test:
+	@$(NX) run research:test
+
+count-lines:
+	@$(NX) run research:stats
+
+all:
+	@$(NX) run paper:build
+
+.PHONY: manim.render manim.render.all
+manim.render:
+	@$(NX) run manim:render
+
+manim.render.all:
+	@$(NX) run manim:render-all

README.md

@@ -1,5 +1,95 @@
+<img width="200" align="left" src="https://github.com/user-attachments/assets/d148b00d-e9f9-4280-89cc-0cc866e17251" />
+
+### PHANTOM
+
+[![Dataset on HF](https://huggingface.co/datasets/huggingface/badges/resolve/main/dataset-on-hf-sm.svg)](https://huggingface.co/datasets/velocitatem/whoclickedit)
 [![Build PDF](https://github.com/velocitatem/PHANTOM/actions/workflows/latex.yml/badge.svg)](https://github.com/velocitatem/PHANTOM/actions/workflows/latex.yml)
+[![Paper](https://img.shields.io/badge/Paper-PDF-red?logo=adobe-acrobat-reader)](https://pub-d5b94a3c29fd40c6b3881946e463fdb7.r2.dev/thesis-latest.pdf)
+[![TPU Research Cloud](https://img.shields.io/badge/TPU%20Research%20Cloud-TRC%20supported-4285F4?logo=googlecloud&logoColor=white)](https://sites.research.google/trc/faq/)
+[![Vercel Deploy](https://deploy-badge.vercel.app/?url=https://phantom-hotel.vercel.app&name=Hotel)](https://phantom-hotel.vercel.app)
+[![Vercel Deploy](https://deploy-badge.vercel.app/?url=https://phantom-airline.vercel.app&name=Airline)](https://phantom-airline.vercel.app)
 
-- https://phantom-hotel.vercel.app/
-- https://phantom-airline.vercel.app/
 
+
+```mermaid
+mindmap
+  PHANTOM((PHANTOM Project))
+    North Star
+      Study how automated actors change markets 
+      Build an experimentation platform for real-world-like commerce 
+      Two-loop learning system
+        Online observation loop 
+        Offline "defense gym" loop 
+    Core Economic Questions
+      Price Discovery
+        How prices respond to demand signals
+        How signal quality changes with bots/agents
+      Demand & Elasticity
+        Shifts in willingness-to-pay
+        Short-run vs long-run elasticity
+      Market Efficiency & Welfare
+        Consumer surplus vs producer surplus
+        Deadweight loss from frictions/manipulation
+      Price Discrimination & Segmentation
+        Behavioral feature-based segmentation
+        Fairness vs profitability tradeoffs
+      Information Asymmetry
+        Agents amplify search and arbitrage
+        Sellers infer more about buyers; buyers infer more about sellers
+      Strategic Interaction
+        Consumers vs firms vs agents
+        Feedback loops: policy ↔ behavior ↔ price
+      Market Power & Competition
+        Algorithmic pricing as competitive tool
+        Risks: tacit coordination / "algorithmic collusion"
+      Externalities
+        Congestion and attention costs
+        Spillovers: one segment’s behavior affects others’ prices
+    System-Level View
+      Participants
+        Humans
+        Agents (automated buyers/actors) 
+        Firms (pricing decision-makers)
+        Platform (measurement + control layer)
+      Markets Simulated
+        Repeated transactions
+        Limited inventory / capacity constraints (conceptually)
+        Time dynamics (learning over time)
+      Interventions
+        Pricing policies
+        Experiment assignment / randomized exposure
+        Agent behavioral policies (task-driven)
+    Measurement & Causal Inference
+      What is observed
+        Actions (search, click, purchase intent)
+        Context (product attributes, time, exposure)
+        Outcomes (conversion, revenue, churn proxies)
+      Identification strategy
+        A/B tests and randomization
+        Counterfactual baselines
+        Robustness checks (offline replay)
+      Key metrics
+        Revenue / profit proxies
+        Conversion & bounce
+        Price volatility / stability
+        Welfare proxies (e.g., dispersion, access)
+    Risk, Governance, and Ethics
+      Manipulation & Integrity
+        Bot-driven demand distortion
+        Measurement contamination
+      Fairness & Transparency
+        Differential pricing concerns
+        Explainability and auditability
+      Safety Constraints
+        Guardrails on price moves
+        Monitoring for runaway feedback loops
+    Outputs
+      Insights
+        When do agents raise/lower prices via behavior shifts?
+        Which market designs are robust to automation?
+      Defenses
+        Agent-aware pricing policies (robust control)
+        Detection + mitigation strategies (feature-level separability)
+      Platform Value
+        Reusable testbed for market + AI-agent research
+```

backend/project.json

@@ -0,0 +1,33 @@
+{
+  "$schema": "../node_modules/nx/schemas/project-schema.json",
+  "name": "platform",
+  "projectType": "application",
+  "sourceRoot": "backend",
+  "targets": {
+    "up": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "docker compose up -d",
+        "cwd": "."
+      }
+    },
+    "down": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "docker compose down",
+        "cwd": "."
+      }
+    },
+    "logs": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "docker compose logs --tail=100 -f",
+        "cwd": "."
+      }
+    }
+  },
+  "tags": [
+    "scope:platform",
+    "type:infra"
+  ]
+}

backend/provider/app.py

@@ -0,0 +1,112 @@
+from fastapi import FastAPI, HTTPException, Query
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel
+from typing import Literal, Optional
+import uvicorn, os, sys
+from supabase import create_client, Client
+from dotenv import load_dotenv
+import numpy as np
+import pandas as pd
+load_dotenv()
+
+# Local imports of registry and pricing function
+
+sys.path.append(os.path.dirname(os.path.abspath(__file__))+ "/../../experiments/")
+from procesing.providers import SupabaseProvider, BackendAPIProvider
+from procesing.pricers import (
+    StaticPricer,
+    RandomPricer,
+    ElasticityBasedPricer
+)
+from procesing.steps import (
+    PredictPricesStep
+)
+from procesing import PipelineContext
+sys.path.append(os.path.dirname(os.path.abspath(__file__))+ "/../../lib/")
+print(os.path.dirname(os.path.abspath(__file__))+ "/../../lib/")
+from lib.model_registry import ModelRegistry
+
+# Config
+app = FastAPI(title="PHANTOM Pricing Provider")
+app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_credentials=True, allow_methods=["*"], allow_headers=["*"])
+
+supabase: Client = create_client(os.getenv("NEXT_PUBLIC_SUPABASE_URL"), os.getenv("NEXT_PUBLIC_SUPABASE_ANON_KEY"))
+registry = ModelRegistry()
+
+class PriceResponse(BaseModel):
+    productId: str
+    price: float
+    base_price: float
+    markup: float
+    elasticity: Optional[float] = None
+    model_version: str = 'latest'
+
+@app.get("/health")
+def health() -> dict:
+    return {"status": "healthy", "redis": registry.health_check()}
+
+@app.get("/api/{mode}/price/{productId}", response_model=PriceResponse)
+def get_price(mode: Literal['hotel', 'airline'], productId: str, sessionId: Optional[str] = Query(None), experimentId: Optional[str] = Query(None)):
+    """
+    THIS is the fast lookup service (mechanism).
+    Priority: session-keyed price > global optimal price > base price
+    """
+    product = supabase.table(f'{mode}_products').select("metadata").eq('id', productId).execute().data[0]
+    if not product: raise HTTPException(404, f"Product {productId} not found")
+
+    metadata = product['metadata']
+    base_price = metadata.get('base_price', 100.0)
+
+    # PRIORITY 1: session-aware price (computed by Airflow worker)
+    if sessionId:
+        session_price = registry.get_session_price(sessionId, productId)
+        if session_price is not None:
+            return PriceResponse(
+                productId=productId,
+                price=session_price,
+                base_price=base_price,
+                markup=session_price/base_price,
+                elasticity=None,
+                model_version='session-aware'
+            )
+
+    # PRIORITY 2: global pre-computed prices (surge pricing)
+    prices_df = registry.get_prices('latest')
+    if prices_df is not None:
+        product_price_row = prices_df[prices_df['productId'] == productId]
+        if not product_price_row.empty:
+            optimal_price = float(product_price_row['optimal_price'].iloc[0])
+            return PriceResponse(
+                productId=productId,
+                price=optimal_price,
+                base_price=base_price,
+                markup=optimal_price/base_price,
+                elasticity=None,
+                model_version='surge'
+            )
+
+    # PRIORITY 3: fallback to base price
+    return PriceResponse(
+        productId=productId,
+        price=base_price,
+        base_price=base_price,
+        markup=1.0,
+        elasticity=None,
+        model_version='base'
+    )
+
+@app.get("/models")
+def list_models(): return registry.list_models()
+
+@app.post("/models/reload")
+def reload_models():
+    elasticity, pricing_model = registry.get_elasticity('latest'), registry.get_pricing_model('latest')
+    return {
+        "elasticity_loaded": bool(elasticity),
+        "n_products": len(elasticity) if elasticity is not None else 0,
+        "pricing_model_loaded": bool(pricing_model),
+        "model_class": pricing_model.__class__.__name__ if pricing_model else None
+    }
+
+if __name__ == "__main__":
+    uvicorn.run(app, host="0.0.0.0", port=int(os.getenv("PROVIDER_PORT", "5001")))

backend/provider/project.json

@@ -0,0 +1,39 @@
+{
+  "$schema": "../../node_modules/nx/schemas/project-schema.json",
+  "name": "pricing-provider",
+  "projectType": "application",
+  "sourceRoot": "backend/provider",
+  "targets": {
+    "install": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash -lc '[ -x ../../.venv/bin/python ] || python3 -m venv ../../.venv; ../../.venv/bin/python -m ensurepip --upgrade; ../../.venv/bin/python -m pip install -r requirements.txt'",
+        "cwd": "backend/provider"
+      }
+    },
+    "dev": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "../../.venv/bin/uvicorn app:app --host 0.0.0.0 --port ${PROVIDER_PORT:-5001} --reload",
+        "cwd": "backend/provider"
+      }
+    },
+    "start": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "../../.venv/bin/uvicorn app:app --host 0.0.0.0 --port ${PROVIDER_PORT:-5001}",
+        "cwd": "backend/provider"
+      }
+    }
+  },
+  "tags": [
+    "scope:backend",
+    "type:provider"
+  ]
+}

backend/provider/requirements.txt

@@ -0,0 +1,16 @@
+fastapi
+uvicorn[standard]
+pydantic
+numpy
+pandas
+scikit-learn
+redis
+supabase
+confluent-kafka>=2.3.0
+kafka-python
+graphviz
+python-dotenv>=1.0.0
+requests>=2.31.0
+typing-extensions>=4.8.0
+pypickle
+pymc

backend/server/app.py

@@ -11,6 +11,7 @@ from kafka import KafkaProducer, KafkaAdminClient, KafkaConsumer
 from kafka.admin import NewTopic
 from kafka.errors import TopicAlreadyExistsError
 from dotenv import load_dotenv
+from supabase import create_client, Client
 load_dotenv()
 
 app = FastAPI()
@@ -18,6 +19,19 @@ app = FastAPI()
 # kafka producer - lazy init
 _producer: Optional[KafkaProducer] = None
 
+# supabase client - lazy init
+_supabase: Optional[Client] = None
+
+def get_supabase() -> Client:
+    global _supabase
+    if _supabase is None:
+        url = os.getenv('NEXT_PUBLIC_SUPABASE_URL')
+        key = os.getenv('NEXT_PUBLIC_SUPABASE_ANON_KEY')
+        if not url or not key:
+            raise ValueError("Supabase credentials not configured")
+        _supabase = create_client(url, key)
+    return _supabase
+
 def get_producer() -> KafkaProducer:
     global _producer
     if _producer is None:
@@ -50,6 +64,14 @@ class EventPayload(BaseModel):
     userAgent: Optional[str] = None
     ts: Optional[str] = None
 
+class PriceLogPayload(BaseModel):
+    productId: str
+    price: float
+    sessionId: str
+    experimentId: Optional[str] = None
+    storeMode: str
+    ts: Optional[str] = None
+
 app.add_middleware(
     CORSMiddleware,
     allow_origins=["*"],
@@ -73,7 +95,8 @@ async def startup_event():
         )
 
         topics = [
-            NewTopic(name='user-interactions', num_partitions=3, replication_factor=1)
+            NewTopic(name='user-interactions', num_partitions=3, replication_factor=1),
+            NewTopic(name='price-logs', num_partitions=3, replication_factor=1)
         ]
 
         admin.create_topics(new_topics=topics, validate_only=False)
@@ -125,42 +148,71 @@ async def ingest_logs(event: EventPayload):
         print(traceback.format_exc())
         raise HTTPException(status_code=500, detail=str(e))
 
+@app.post("/api/kafka/price-log")
+async def ingest_price_log(price_log: PriceLogPayload):
+    try:
+        if not price_log.ts:
+            price_log.ts = datetime.utcnow().isoformat() + 'Z'
+
+        producer = get_producer()
+        future = producer.send(
+            'price-logs',
+            key=price_log.productId,
+            value=price_log.model_dump()
+        )
+        future.add_errback(lambda e: print(f"[KAFKA_PRICE_LOG_ERROR] {e}"))
+
+        return {"success": True}
+    except Exception as e:
+        import traceback
+        print(f"[PRICE_LOG_ERROR] {e}")
+        print(traceback.format_exc())
+        raise HTTPException(status_code=500, detail=str(e))
+
 @app.get("/api/kafka/dump")
 def dump_logs(
+    topic: str = 'user-interactions',
     last_n: Optional[int] = None,
     t_start: Optional[str] = None,
     t_end: Optional[str] = None
 ):
-    """dump all messages from user-interactions topic
+    """dump all messages from specified kafka topic
 
     params:
+        topic: kafka topic to dump (default: user-interactions)
         last_n: return only last n messages (default: all)
-        t_start: filter by start timestamp iso format (future use)
-        t_end: filter by end timestamp iso format (future use)
+        t_start: filter by start timestamp iso format
+        t_end: filter by end timestamp iso format
     """
+    if topic not in ['user-interactions', 'price-logs']:
+        raise HTTPException(status_code=400, detail="Invalid topic")
+
     host = os.getenv('KAFKA_HOST', 'localhost')
     port = os.getenv('KAFKA_PORT', '9092')
     broker = f'{host}:{port}'
 
     try:
         consumer = KafkaConsumer(
-            'user-interactions',
+            topic,
             bootstrap_servers=[broker],
             auto_offset_reset='earliest',
             enable_auto_commit=False,
             value_deserializer=lambda x: json.loads(x.decode('utf-8')),
-            consumer_timeout_ms=5000
+            consumer_timeout_ms=30000,
+            fetch_max_wait_ms=10000,
+            max_poll_records=1000
         )
 
         events = []
         for msg in consumer:
             events.append(msg.value)
+            if last_n and len(events) >= last_n * 2:
+                break
 
         consumer.close()
 
         # apply filters
         if t_start or t_end:
-            # filter by timestamp range if provided
             filtered = []
             for e in events:
                 ts = e.get('ts')
@@ -183,6 +235,131 @@ def dump_logs(
         print(traceback.format_exc())
         raise HTTPException(status_code=500, detail=str(e))
 
+@app.get("/api/products/{product_id}")
+async def get_product_by_id(product_id: str):
+    """fetch single product by id from either hotel_products or airline_products"""
+    try:
+        supabase = get_supabase()
+
+        # try hotel_products first
+        response = supabase.table('hotel_products').select('*').eq('id', product_id).execute()
+        if response.data and len(response.data) > 0:
+            return {"success": True, "data": response.data[0]}
+
+        # try airline_products
+        response = supabase.table('airline_products').select('*').eq('id', product_id).execute()
+        if response.data and len(response.data) > 0:
+            return {"success": True, "data": response.data[0]}
+
+        raise HTTPException(status_code=404, detail="Product not found")
+
+    except HTTPException:
+        raise
+    except Exception as e:
+        import traceback
+        print(f"[PRODUCT_BY_ID_ERROR] {e}")
+        print(traceback.format_exc())
+        raise HTTPException(status_code=500, detail=str(e))
+
+@app.get("/api/products/type/{product_type}")
+async def get_products(
+    product_type: str,
+    dateIndex: Optional[int] = None,
+    origin: Optional[str] = None,
+    destination: Optional[str] = None,
+    tripType: Optional[str] = None,
+    adults: Optional[int] = None,
+    children: Optional[int] = None,
+    infants: Optional[int] = None,
+    rooms: Optional[int] = None
+):
+    """fetch products from supabase based on type (hotel or airline)
+
+    params:
+        product_type: either 'hotel' or 'airline'
+        dateIndex: optional days offset from today (e.g., 0=today, 1=tomorrow, -1=yesterday)
+        origin: (airline) departure airport code
+        destination: (airline/hotel) arrival airport or hotel location
+        tripType: (airline) roundtrip, oneway, multicity
+        adults, children, infants: passenger counts
+        rooms: (hotel) number of rooms
+    """
+    if product_type not in ['hotel', 'airline']:
+        raise HTTPException(status_code=400, detail="product_type must be 'hotel' or 'airline'")
+
+    try:
+        supabase = get_supabase()
+        table = f'{product_type}_products'
+
+        query = supabase.table(table).select('*')
+
+        # filter by exact date_index if provided
+        # dateIndex from frontend is days from today, convert to days since epoch
+        if dateIndex is not None:
+            query = query.eq('date_index', dateIndex)
+
+        response = query.execute()
+        results = response.data
+
+        # apply in-memory filters based on metadata for airline products
+        if product_type == 'airline' and results:
+            filtered = []
+            for product in results:
+                metadata = product.get('metadata', {})
+
+                # filter by origin airport
+                if origin:
+                    dep = metadata.get('departure', {})
+                    if dep.get('airport') != origin:
+                        continue
+
+                # filter by destination airport
+                if destination:
+                    arr = metadata.get('arrival', {})
+                    if arr.get('airport') != destination:
+                        continue
+
+                # passenger count validation (ensure total capacity)
+                if adults is not None or children is not None or infants is not None:
+                    total_pax = (adults or 0) + (children or 0) + (infants or 0)
+                    avail = product.get('availability', 0)
+                    if avail < total_pax:
+                        continue
+
+                filtered.append(product)
+
+            results = filtered
+
+        # apply in-memory filters for hotel products
+        elif product_type == 'hotel' and results:
+            filtered = []
+            for product in results:
+                metadata = product.get('metadata', {})
+
+                # filter by occupancy capacity
+                if adults is not None:
+                    max_occ = metadata.get('max_occupancy', 2)
+                    if max_occ < adults:
+                        continue
+
+                # filter by room availability
+                if rooms is not None:
+                    avail = product.get('availability', 0)
+                    if avail < rooms:
+                        continue
+
+                filtered.append(product)
+
+            results = filtered
+
+        return {"success": True, "count": len(results), "data": results}
+
+    except Exception as e:
+        import traceback
+        print(f"[PRODUCTS_ERROR] {e}")
+        print(traceback.format_exc())
+        raise HTTPException(status_code=500, detail=str(e))
+
 
 
 if __name__ == "__main__":

backend/server/project.json

@@ -0,0 +1,39 @@
+{
+  "$schema": "../../node_modules/nx/schemas/project-schema.json",
+  "name": "backend-server",
+  "projectType": "application",
+  "sourceRoot": "backend/server",
+  "targets": {
+    "install": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash -lc '[ -x ../../.venv/bin/python ] || python3 -m venv ../../.venv; ../../.venv/bin/python -m ensurepip --upgrade; ../../.venv/bin/python -m pip install -r requirements.txt'",
+        "cwd": "backend/server"
+      }
+    },
+    "dev": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "../../.venv/bin/uvicorn app:app --host 0.0.0.0 --port ${BACKEND_PORT:-5000} --reload",
+        "cwd": "backend/server"
+      }
+    },
+    "start": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "../../.venv/bin/uvicorn app:app --host 0.0.0.0 --port ${BACKEND_PORT:-5000}",
+        "cwd": "backend/server"
+      }
+    }
+  },
+  "tags": [
+    "scope:backend",
+    "type:api"
+  ]
+}

backend/server/requirements.txt

@@ -1,5 +1,6 @@
-fastapi==0.104.1
-uvicorn[standard]==0.24.0
-kafka-python==2.0.2
-pydantic==2.5.0
-python-dotenv==1.0.0
+fastapi>=0.135,<0.136
+uvicorn[standard]>=0.41,<0.42
+kafka-python>=2.3,<2.4
+pydantic>=2.12,<3
+python-dotenv>=1.0,<2
+supabase>=2.28,<3

backend/worker/project.json

@@ -0,0 +1,39 @@
+{
+  "$schema": "../../node_modules/nx/schemas/project-schema.json",
+  "name": "backend-worker",
+  "projectType": "application",
+  "sourceRoot": "backend/worker",
+  "targets": {
+    "install": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash -lc '[ -x ../../.venv/bin/python ] || python3 -m venv ../../.venv; ../../.venv/bin/python -m ensurepip --upgrade; ../../.venv/bin/python -m pip install -r requirements.txt'",
+        "cwd": "backend/worker"
+      }
+    },
+    "dev": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "../../.venv/bin/celery -A main:app worker --loglevel=info",
+        "cwd": "backend/worker"
+      }
+    },
+    "start": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "../../.venv/bin/python main.py",
+        "cwd": "backend/worker"
+      }
+    }
+  },
+  "tags": [
+    "scope:backend",
+    "type:worker"
+  ]
+}

backend/worker/requirements.txt

@@ -0,0 +1,3 @@
+celery>=5.3,<6
+python-dotenv>=1.0.0
+redis>=5.0.0

docker-compose.yml

@@ -1,4 +1,43 @@
 services:
+  tpu-watchdogs:
+    build:
+      context: .
+      dockerfile: docker/TPUWatchdog.dockerfile
+    container_name: "PHANTOM-tpu-watchdogs"
+    restart: unless-stopped
+    user: "${UID:-1000}:${GID:-1000}"
+    environment:
+      - HF_TOKEN=${HF_TOKEN}
+      - WANDB_API_KEY=${WANDB_API_KEY}
+      - GITHUB_TOKEN=${GITHUB_TOKEN}
+      - GOOGLE_APPLICATION_CREDENTIALS=/secrets/gcp-sa.json
+      - GCP_ACCOUNT=${GCP_ACCOUNT:-}
+      - WATCHDOG_CONFIG_PATTERN=${WATCHDOG_CONFIG_PATTERN:-v[46]*.conf}
+      - CLOUDSDK_CONFIG=/.config/gcloud
+    volumes:
+      - ~/.config/gcloud:/.config/gcloud:rw
+      - ./secrets/gcp-sa.json:/secrets/gcp-sa.json:ro
+
+  tensorboard-rl:
+    image: tensorflow/tensorflow:latest
+    container_name: "PHANTOM-tensorboard-rl"
+    ports:
+      - "6007:6006"
+    volumes:
+      - ./sim/rl/runs:/logs
+    command: tensorboard --logdir=/logs --host=0.0.0.0 --port=6006
+    restart: unless-stopped
+
+  tensorboard-ml:
+    image: tensorflow/tensorflow:latest
+    container_name: "PHANTOM-tensorboard-ml"
+    ports:
+      - "6006:6006"
+    volumes:
+      - ./experiments/ml/runs:/logs
+    command: tensorboard --logdir=/logs --host=0.0.0.0 --port=6006
+    restart: unless-stopped
+
   backend:
     container_name: "PHANTOM-backend"
     build:
@@ -9,6 +48,9 @@ services:
     environment:
       - KAFKA_HOST=kafka
       - KAFKA_PORT=29092
+      - BACKEND_PORT=5000
+      - NEXT_PUBLIC_SUPABASE_URL=${NEXT_PUBLIC_SUPABASE_URL}
+      - NEXT_PUBLIC_SUPABASE_ANON_KEY=${NEXT_PUBLIC_SUPABASE_ANON_KEY}
     depends_on:
       - kafka
     restart: unless-stopped
@@ -68,6 +110,149 @@ services:
       - "${REDPANDA_CONSOLE_PORT:-8080}:8080"
     restart: unless-stopped
 
+  postgres:
+    container_name: "PHANTOM-postgres"
+    image: postgres:13
+    environment:
+      - POSTGRES_USER=airflow
+      - POSTGRES_PASSWORD=airflow
+      - POSTGRES_DB=airflow
+    ports:
+      - "5433:5432"
+    volumes:
+      - postgres_data:/var/lib/postgresql/data
+    restart: unless-stopped
+
+  airflow-init:
+    container_name: "PHANTOM-airflow-init"
+    build:
+      context: .
+      dockerfile: docker/Airflow.dockerfile
+    depends_on:
+      - postgres
+    environment:
+      - AIRFLOW__CORE__EXECUTOR=LocalExecutor
+      - AIRFLOW__DATABASE__SQL_ALCHEMY_CONN=postgresql+psycopg2://airflow:airflow@postgres/airflow
+      - AIRFLOW__CORE__FERNET_KEY=${AIRFLOW_FERNET_KEY}
+      - AIRFLOW__CORE__LOAD_EXAMPLES=false
+      - AIRFLOW__CORE__ENABLE_XCOM_PICKLING=true
+      - AIRFLOW__CORE__PARALLELISM=16
+      - AIRFLOW__CORE__DAG_CONCURRENCY=8
+      - AIRFLOW__CORE__MAX_ACTIVE_RUNS_PER_DAG=4
+      - _AIRFLOW_DB_MIGRATE=true
+      - _AIRFLOW_WWW_USER_CREATE=true
+      - _AIRFLOW_WWW_USER_USERNAME=admin
+      - _AIRFLOW_WWW_USER_PASSWORD=admin
+      - REDIS_HOST=redis
+      - REDIS_PORT=6379
+    command: version
+    restart: "no"
+
+  airflow-webserver:
+    container_name: "PHANTOM-airflow-webserver"
+    build:
+      context: .
+      dockerfile: docker/Airflow.dockerfile
+    depends_on:
+      - postgres
+      - airflow-init
+      - redis
+    environment:
+      - AIRFLOW__CORE__EXECUTOR=LocalExecutor
+      - AIRFLOW__DATABASE__SQL_ALCHEMY_CONN=postgresql+psycopg2://airflow:airflow@postgres/airflow
+      - AIRFLOW__CORE__FERNET_KEY=${AIRFLOW_FERNET_KEY}
+      - AIRFLOW__CORE__DAGS_ARE_PAUSED_AT_CREATION=true
+      - AIRFLOW__CORE__LOAD_EXAMPLES=false
+      - AIRFLOW__CORE__ENABLE_XCOM_PICKLING=true
+      - AIRFLOW__CORE__PARALLELISM=16
+      - AIRFLOW__CORE__DAG_CONCURRENCY=8
+      - AIRFLOW__CORE__MAX_ACTIVE_RUNS_PER_DAG=4
+      - AIRFLOW__SCHEDULER__MIN_FILE_PROCESS_INTERVAL=30
+      - AIRFLOW__SCHEDULER__DAG_DIR_LIST_INTERVAL=60
+      - AIRFLOW__WEBSERVER__EXPOSE_CONFIG=true
+      - AIRFLOW__WEBSERVER__SECRET_KEY=${AIRFLOW_SECRET_KEY}
+      - AIRFLOW__API__AUTH_BACKENDS=airflow.api.auth.backend.basic_auth
+      - KAFKA_HOST=kafka
+      - KAFKA_PORT=29092
+      - BACKEND_URL=http://backend:5000
+      - NEXT_PUBLIC_SUPABASE_URL=${NEXT_PUBLIC_SUPABASE_URL}
+      - NEXT_PUBLIC_SUPABASE_ANON_KEY=${NEXT_PUBLIC_SUPABASE_ANON_KEY}
+      - REDIS_HOST=redis
+      - REDIS_PORT=6379
+    ports:
+      - "${AIRFLOW_WEBSERVER_PORT:-8085}:8080"
+    command: webserver
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "curl", "--fail", "http://localhost:8080/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 5
+      start_period: 30s
+
+  airflow-scheduler:
+    container_name: "PHANTOM-airflow-scheduler"
+    build:
+      context: .
+      dockerfile: docker/Airflow.dockerfile
+    depends_on:
+      airflow-webserver:
+        condition: service_healthy
+      redis:
+        condition: service_started
+    environment:
+      - AIRFLOW__CORE__EXECUTOR=LocalExecutor
+      - AIRFLOW__DATABASE__SQL_ALCHEMY_CONN=postgresql+psycopg2://airflow:airflow@postgres/airflow
+      - AIRFLOW__CORE__FERNET_KEY=${AIRFLOW_FERNET_KEY}
+      - AIRFLOW__CORE__DAGS_ARE_PAUSED_AT_CREATION=true
+      - AIRFLOW__CORE__LOAD_EXAMPLES=false
+      - AIRFLOW__CORE__ENABLE_XCOM_PICKLING=true
+      - AIRFLOW__CORE__PARALLELISM=16
+      - AIRFLOW__CORE__DAG_CONCURRENCY=8
+      - AIRFLOW__CORE__MAX_ACTIVE_RUNS_PER_DAG=4
+      - AIRFLOW__SCHEDULER__MIN_FILE_PROCESS_INTERVAL=30
+      - AIRFLOW__SCHEDULER__DAG_DIR_LIST_INTERVAL=60
+      - AIRFLOW__SCHEDULER__PARSING_PROCESSES=2
+      - AIRFLOW__WEBSERVER__SECRET_KEY=${AIRFLOW_SECRET_KEY}
+      - AIRFLOW__API__AUTH_BACKENDS=airflow.api.auth.backend.basic_auth
+      - KAFKA_HOST=kafka
+      - KAFKA_PORT=29092
+      - BACKEND_URL=http://backend:5000
+      - NEXT_PUBLIC_SUPABASE_URL=${NEXT_PUBLIC_SUPABASE_URL}
+      - NEXT_PUBLIC_SUPABASE_ANON_KEY=${NEXT_PUBLIC_SUPABASE_ANON_KEY}
+      - REDIS_HOST=redis
+      - REDIS_PORT=6379
+    command: scheduler
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD-SHELL", 'airflow jobs check --job-type SchedulerJob --hostname "$${HOSTNAME}"']
+      interval: 30s
+      timeout: 10s
+      retries: 5
+      start_period: 30s
+
+  pricing-provider:
+    container_name: "PHANTOM-pricing-provider"
+    build:
+      context: .
+      dockerfile: docker/Provider.dockerfile
+    depends_on:
+      - redis
+      - kafka
+    environment:
+      - PROVIDER_PORT=5001
+      - REDIS_HOST=redis
+      - REDIS_PORT=6379
+      - KAFKA_HOST=kafka
+      - KAFKA_PORT=29092
+      - NEXT_PUBLIC_SUPABASE_URL=${NEXT_PUBLIC_SUPABASE_URL}
+      - NEXT_PUBLIC_SUPABASE_ANON_KEY=${NEXT_PUBLIC_SUPABASE_ANON_KEY}
+      - BACKEND_URL=http://localhost:5000
+    ports:
+      - "${PROVIDER_PORT:-5001}:5001"
+    restart: unless-stopped
+
 volumes:
   phantom_kafka_data:
   phantom_redis_data:
+  postgres_data:

docker/Airflow.dockerfile

@@ -0,0 +1,30 @@
+FROM apache/airflow:2.7.3-python3.11
+
+USER root
+
+# install system deps if needed
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential \
+    && apt-get clean \
+    && rm -rf /var/lib/apt/lists/*
+
+USER airflow
+
+# copy requirements for pipeline dependencies
+COPY requirements.txt /tmp/requirements.txt
+RUN pip install --no-cache-dir -r /tmp/requirements.txt
+
+# install postgres driver and providers
+RUN pip install --no-cache-dir \
+    psycopg2-binary \
+    apache-airflow-providers-postgres
+
+# set airflow home
+ENV AIRFLOW_HOME=/opt/airflow
+
+COPY --chown=airflow:root experiments/airflow/dags ${AIRFLOW_HOME}/dags
+COPY --chown=airflow:root experiments/procesing ${AIRFLOW_HOME}/procesing
+COPY --chown=airflow:root lib ${AIRFLOW_HOME}/lib
+
+# create logs and plugins dirs (airflow expects them)
+RUN mkdir -p ${AIRFLOW_HOME}/logs ${AIRFLOW_HOME}/plugins

docker/Airflow.railway.dockerfile

@@ -0,0 +1,41 @@
+FROM apache/airflow:2.7.3-python3.11
+
+USER root
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential \
+    supervisor \
+    && apt-get clean \
+    && rm -rf /var/lib/apt/lists/*
+
+USER airflow
+
+COPY requirements.txt /tmp/requirements.txt
+RUN pip install --no-cache-dir -r /tmp/requirements.txt
+
+RUN pip install --no-cache-dir \
+    psycopg2-binary \
+    apache-airflow-providers-postgres
+
+ENV AIRFLOW_HOME=/opt/airflow
+ENV AIRFLOW__CORE__EXECUTOR=SequentialExecutor
+ENV AIRFLOW__CORE__LOAD_EXAMPLES=false
+ENV AIRFLOW__CORE__ENABLE_XCOM_PICKLING=true
+ENV AIRFLOW__WEBSERVER__EXPOSE_CONFIG=true
+
+# copy all code into image (standalone - no volume mounts needed)
+COPY --chown=airflow:root experiments/airflow/dags ${AIRFLOW_HOME}/dags
+COPY --chown=airflow:root experiments/procesing ${AIRFLOW_HOME}/procesing
+COPY --chown=airflow:root lib ${AIRFLOW_HOME}/lib
+
+RUN mkdir -p ${AIRFLOW_HOME}/logs ${AIRFLOW_HOME}/plugins
+
+# copy entrypoint script
+COPY --chown=airflow:root docker/airflow-railway-entrypoint.sh /entrypoint.sh
+USER root
+RUN chmod +x /entrypoint.sh
+USER airflow
+
+EXPOSE 8080
+
+ENTRYPOINT ["/entrypoint.sh"]

docker/Provider.dockerfile

@@ -0,0 +1,26 @@
+FROM python:3.11-slim
+
+WORKDIR /app
+
+# Install system dependencies including graphviz
+RUN apt-get update && apt-get install -y \
+    gcc \
+    g++ \
+    graphviz \
+    libgraphviz-dev \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copy and install Python dependencies
+COPY backend/provider/requirements.txt /app/
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Copy application code into image
+COPY lib/ /app/lib/
+COPY experiments/procesing/ /app/procesing/
+COPY backend/provider/ /app/provider/
+
+ENV PYTHONPATH=/app:/app/lib:/app/procesing
+
+WORKDIR /app/provider
+
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "5001"]

docker/TPUWatchdog.dockerfile

@@ -0,0 +1,112 @@
+FROM google/cloud-sdk:slim
+
+# Install tmux to manage multiple watchdogs and jq for json parsing
+RUN apt-get update && \
+    apt-get install -y tmux jq && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+# Copy the orchestration scripts and configs
+COPY tpu_orchestration/ /app/tpu_orchestration/
+
+# Make sure scripts are executable
+RUN chmod +x /app/tpu_orchestration/watchdog.sh
+RUN chmod +x /app/tpu_orchestration/tpu_startup.sh
+
+# Create an entrypoint script that launches a watchdog for each config
+COPY <<-'EOF' /app/entrypoint.sh
+#!/bin/bash
+set -e
+
+# Make sure required variables are set
+if [ -z "$HF_TOKEN" ]; then
+    echo "Error: HF_TOKEN environment variable is required."
+    exit 1
+fi
+
+if [ -z "$WANDB_API_KEY" ]; then
+    echo "Warning: WANDB_API_KEY environment variable is not set. Wandb logging may fail on TPUs."
+fi
+
+# Authenticate gcloud if credentials are provided
+if [ -n "$GOOGLE_APPLICATION_CREDENTIALS" ] && [ -f "$GOOGLE_APPLICATION_CREDENTIALS" ]; then
+    CRED_TYPE=$(jq -r '.type' "$GOOGLE_APPLICATION_CREDENTIALS" 2>/dev/null || echo "unknown")
+    if [ "$CRED_TYPE" = "service_account" ]; then
+        echo "Authenticating gcloud using service account key..."
+        gcloud auth activate-service-account --key-file="$GOOGLE_APPLICATION_CREDENTIALS"
+
+        if [ -z "$PROJECT_ID" ]; then
+            PROJECT_ID=$(jq -r '.project_id // empty' "$GOOGLE_APPLICATION_CREDENTIALS")
+        fi
+    elif [ "$CRED_TYPE" = "authorized_user" ]; then
+        echo "Using authorized_user credentials via credential file override..."
+        export CLOUDSDK_AUTH_CREDENTIAL_FILE_OVERRIDE="$GOOGLE_APPLICATION_CREDENTIALS"
+
+        if gcloud auth print-access-token >/dev/null 2>&1; then
+            ACTIVE_ACCOUNT=$(gcloud config get-value account 2>/dev/null || true)
+            if [ -z "$ACTIVE_ACCOUNT" ] || [ "$ACTIVE_ACCOUNT" = "(unset)" ]; then
+                ACTIVE_ACCOUNT=$(jq -r '.account // empty' "$GOOGLE_APPLICATION_CREDENTIALS")
+            fi
+
+            if [ -n "$ACTIVE_ACCOUNT" ] && [ "$ACTIVE_ACCOUNT" != "(unset)" ]; then
+                echo "Using gcloud account: $ACTIVE_ACCOUNT"
+            else
+                echo "Using gcloud credential override from $GOOGLE_APPLICATION_CREDENTIALS"
+            fi
+        else
+            echo "Warning: credential file override token check failed. Falling back to mounted gcloud config."
+            unset CLOUDSDK_AUTH_CREDENTIAL_FILE_OVERRIDE
+
+            if [ -n "$GCP_ACCOUNT" ]; then
+                gcloud config set account "$GCP_ACCOUNT" >/dev/null 2>&1 || true
+            fi
+
+            ACTIVE_ACCOUNT=$(gcloud config get-value account 2>/dev/null || true)
+            if [ -z "$ACTIVE_ACCOUNT" ] || [ "$ACTIVE_ACCOUNT" = "(unset)" ]; then
+                echo "Error: no active gcloud account available. Run 'gcloud auth login' on host and mount ~/.config/gcloud, or use a service account key."
+                exit 1
+            fi
+            echo "Using gcloud account: $ACTIVE_ACCOUNT"
+        fi
+    else
+        echo "Warning: unsupported credential file type '$CRED_TYPE'. Falling back to mounted gcloud config."
+    fi
+else
+    echo "Note: Assuming gcloud config is mounted from host."
+fi
+
+if [ -n "$PROJECT_ID" ]; then
+    gcloud config set project "$PROJECT_ID"
+    echo "Set project to $PROJECT_ID"
+fi
+
+# Run the watchdogs in the background using bash instead of tmux
+# Tmux needs a TTY to attach properly which we might not have in docker
+# Stagger startups by 15s to prevent simultaneous TPU creation quota hits
+CONFIG_PATTERN=${WATCHDOG_CONFIG_PATTERN:-"*.conf"}
+shopt -s nullglob
+CONFIGS=(/app/tpu_orchestration/configs/$CONFIG_PATTERN)
+
+if [ ${#CONFIGS[@]} -eq 0 ]; then
+    echo "Error: no watchdog configs matched pattern '$CONFIG_PATTERN'."
+    exit 1
+fi
+
+echo "Using watchdog config pattern: $CONFIG_PATTERN"
+DELAY=0
+for conf in "${CONFIGS[@]}"; do
+    echo "Starting watchdog for $(basename "$conf" .conf) (delay: ${DELAY}s)"
+    (sleep $DELAY && /app/tpu_orchestration/watchdog.sh "$conf") &
+    DELAY=$((DELAY + 15))
+done
+
+echo "All watchdogs queued with staggered startup."
+
+# Keep the container running
+wait
+EOF
+
+RUN chmod +x /app/entrypoint.sh
+
+CMD ["/app/entrypoint.sh"]

docker/Trainer.dockerfile

@@ -0,0 +1,15 @@
+# syntax=docker/dockerfile:1.7
+
+FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime AS gpu
+
+WORKDIR /app
+
+COPY docker/trainer.requirements.txt /tmp/requirements.txt
+RUN pip install --no-cache-dir -r /tmp/requirements.txt
+
+COPY --chmod=755 docker/trainer-agent-entrypoint.sh /usr/local/bin/trainer-agent-entrypoint
+COPY engine /app/engine
+
+ENV PYTHONPATH=/app
+
+ENTRYPOINT ["/usr/local/bin/trainer-agent-entrypoint"]

docker/airflow-railway-entrypoint.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+set -e
+
+# init db and create admin user on first run
+airflow db migrate
+
+# create admin user if not exists
+airflow users create \
+    --username "${AIRFLOW_ADMIN_USER:-admin}" \
+    --password "${AIRFLOW_ADMIN_PASSWORD:-admin}" \
+    --firstname Admin \
+    --lastname User \
+    --role Admin \
+    --email admin@example.com || true
+
+# start scheduler in background
+airflow scheduler &
+
+# start webserver in foreground (Railway needs one foreground process)
+exec airflow webserver --port ${PORT:-8080}

docker/trainer-agent-entrypoint.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env sh
+set -eu
+
+if [ -z "${SWEEP_ID:-}" ]; then
+  echo "SWEEP_ID is required"
+  exit 1
+fi
+
+set -- python -m engine.train --sweep-agent --sweep-id "${SWEEP_ID}"
+
+if [ -n "${PHANTOM_DEFAULT_AGENT_ARGS:-}" ]; then
+  set -- "$@" ${PHANTOM_DEFAULT_AGENT_ARGS}
+fi
+
+if [ -n "${TRAIN_ARGS:-}" ]; then
+  set -- "$@" ${TRAIN_ARGS}
+fi
+
+if [ "${AGENT_COUNT:-0}" != "0" ]; then
+  set -- "$@" --count "${AGENT_COUNT}"
+fi
+
+exec "$@"

docker/trainer.requirements.txt

@@ -0,0 +1,7 @@
+numpy>=1.24.0
+pandas>=2.0.0
+scipy>=1.11.0
+gymnasium>=0.29.0
+stable-baselines3>=2.2.0
+tensorboard>=2.15.0
+wandb>=0.17.0

docs/goals/goals.csv

@@ -0,0 +1,21 @@
+store_mode,task_name,task_description,definition_of_done
+airline,The Indecisive Executive (SEA-LAX),"You are traveling SEA to LAX for business. You prefer Business Class for the comfort, but you need to justify the expense to your company. 1) Find the Business Class option and check its price. 2) Compare it against the Economy option on the same route to see how much money you are saving or spending. 3) Spend some time weighing the pros and cons of the ""Flexible"" fare rule vs the standard one. 4) Ultimately, decide that your comfort is worth it and book the Business Class ticket.","Booking for SEA-LAX Business Class is completed."
+airline,The Cross-Country Splurge (LAX-JFK),"You are flying LAX to JFK and want to treat yourself to First Class, but only if it's the right flight. 1) Find the First Class option. 2) thoroughly check the details (duration, arrival time). 3) Compare it with the Business Class option if available, or just look at other departure times to ensure this is the best schedule. 4) After confirming this is the absolute best option, proceed to book First Class.","Booking for LAX-JFK First Class is completed."
+airline,The Budget Student (DFW-ORD),"You are a broke student flying DFW to ORD. You have a budget of roughly $200. 1) Find the cheapest Economy flight. 2) Before booking, frantically check if there are any other flights or if the ""Premium"" economy is somehow cheaper (it won't be, but you should check). 3) Hesitate for a moment to consider if you should just drive instead. 4) Resign yourself to the flight and book the Economy ticket.","Booking for DFW-ORD Economy Class is completed."
+airline,The Quick Hop Commuter (LAX-SFO),"You need to get from LAX to SFO as fast as possible. Price is secondary to speed. 1) Search for flights and identify the one with the shortest duration (1h 30m). 2) Click into the details to verify the arrival time fits your schedule. 3) briefly explore if there's a Business Class upgrade available for this short flight. 4) Decide to stick with Economy since it's such a short trip and book it.","Booking for LAX-SFO is completed."
+airline,The Status Chaser (SFO-SEA),"You are trying to earn airline points and need a ""Premium"" class ticket specifically. 1) Search SFO to SEA. 2) Filter or look for the Premium Economy option. 3) Compare the price gap between Premium and Standard Economy. 4) Browse the details to see if the ""Premium"" fare includes better baggage allowance. 5) Conclude it's worth the points and book the Premium seat.","Booking for SFO-SEA Premium Economy is completed."
+airline,The Family Reunion (MIA-ATL),"You are booking for a family of 4 (2 adults, 2 children) flying MIA to ATL. 1) Search for 4 passengers. 2) You prefer Premium, but if the total is too high, you might settle for Economy. 3) Add Premium to your cart, look at the total, and hesitate. 4) Go back and check the Economy price for 4 people. 5) Decide to treat your family and go back to book the Premium option.","Booking for MIA-ATL (Premium) is completed."
+airline,The Red Eye Skeptic (LAX-JFK),"You need to fly LAX to JFK but hate late arrivals. 1) Search for the flight and check the arrival time of the First Class option. 2) It arrives early morning (02:15), which worries you. 3) Spend some time looking for other flight options on different days to see if there's a better schedule. 4) Realize this is the only direct option that works and proceed to book it despite the time.","Booking for LAX-JFK is completed."
+airline,The Refundable Requirement (ATL-DFW),"Your meeting in Dallas might get cancelled, so you strictly need a ""Refundable"" ticket. 1) Search ATL to DFW. 2) Find the First Class option and verify it lists ""Refundable"". 3) Check the Economy option to see if it is also refundable (it might not be). 4) Weigh the cost difference. 5) Choose the First Class Refundable option for peace of mind.","Booking for ATL-DFW First Class is completed."
+airline,The Hub Connector (ORD-MIA),"You are flying ORD to MIA to catch a cruise. You cannot be late. 1) Search for the flight. 2) Verify the ""stops"" is 0 (Direct). 3) Click into details to check the duration. 4) Worry that 3h 30m might be too long in Economy. 5) Look for a Business class option. 6) Decide to save money for the cruise and book Economy.","Booking for ORD-MIA Economy is completed."
+airline,The West Coast Hopper (SEA-LAX Business),"You fly this route often and usually pay around $700. 1) Search SEA to LAX. 2) Find the Business Class ticket. 3) Check if the price is near your usual $720 or if it's surged. 4) If it looks expensive, browse other dates to compare. 5) Return to your original desired date and book the Business Class seat.","Booking for SEA-LAX Business is completed."
+hotel,The Honeymoon Suite (Presidential),"It is your honeymoon. You want the best room available, specifically one with a ""jacuzzi"". 1) Search for a room for 2 people. 2) Identify the ""Presidential Suite"". 3) Click details to confirm the amenities include a jacuzzi. 4) Browse the ""Executive Suite"" just to see what you are upgrading from. 5) Go back to the Presidential Suite, confirm it's the one you want, and book it.","Booking for the Presidential Suite is completed."
+hotel,The Digital Nomad (Executive),"You are working remotely and strictly need a ""workspace"". 1) Search for a room. 2) Check the ""Executive Suite"" details for a workspace. 3) Check the ""Deluxe Room"" to see if it also has a workspace and is cheaper. 4) Compare the images (if available) or amenity lists of both. 5) Decide the Executive Suite looks more comfortable for a week of work and book it.","Booking for the Executive Suite is completed."
+hotel,The Safety First (Superior),"You are traveling with valuables and need a ""safe"" in the room. 1) Search for a room. 2) Look at the ""Standard Room"" amenities. Does it have a safe? 3) Look at the ""Superior Room"". Verify it has a safe. 4) Compare the price difference. Is safety worth the extra cost? 5) Decide it is, and book the Superior Room.","Booking for the Superior Room is completed."
+hotel,The Bachelor Party (Max Occupancy),"You are booking for 4 guys. You want everyone in one room if possible. 1) Search for 4 adults. 2) Find the room that fits 4 people (Presidential). 3) It looks expensive. Go back and search for 2 adults to see the price of a ""Standard Room"". 4) Calculate if booking two Standard Rooms is cheaper than one Presidential. 5) Decide it's too much hassle to manage two bookings and book the Presidential Suite.","Booking for the Presidential Suite is completed."
+hotel,The Budget Refundable (Junior),"You want a cheap room but your dates might change, so it MUST be refundable. 1) Search for a room. 2) Sort by price or find the cheapest options. 3) Check the ""Standard"" and ""Superior"" rooms. Notice they are likely Non-Refundable. 4) Find the ""Junior Suite"" which is Refundable. 5) Grumble about the price difference but book the Junior Suite because you need the flexibility.","Booking for the Junior Suite is completed."
+hotel,The View Hunter (Executive),"You want a room with a ""city_view"" or balcony. 1) Search for a room. 2) Check the amenities of the ""Deluxe Room"". 3) Check the amenities of the ""Executive Suite"". 4) Compare the prices. 5) Decide to treat yourself to the Executive Suite for the better view/balcony and book it.","Booking for the Executive Suite is completed."
+hotel,The Just-A-Bed (Standard),"You just need a place to crash. Lowest price wins. 1) Search for a room. 2) Identify the absolute cheapest option (Standard Room). 3) Click details just to make sure it has ""wifi"". 4) Briefly glance at the ""Superior Room"" to see if the upgrade is <$10. 5) If not, go back and book the Standard Room immediately.","Booking for the Standard Room is completed."
+hotel,The Family Vacation (Deluxe),"You are traveling with a child. You need a room that isn't too cramped but not a suite. 1) Search for 2 adults, 1 child. 2) Look at the ""Deluxe Room"". 3) Check the amenities for ""coffee_maker"" (parents need coffee). 4) Compare it with the ""Junior Suite"". 5) Decide the Deluxe Room is sufficient value and book it.","Booking for the Deluxe Room is completed."
+hotel,The Long Stay (Junior),"You are staying for 7 nights. You want something nicer than a standard room but affordable. 1) Search for a room. 2) Look at the ""Junior Suite"". 3) Check the amenities for a ""mini_fridge"" or similar. 4) Compare the total cost for 7 nights against your budget. 5) Hesitate and look at the ""Standard Room"" price. 6) Decide the extra space of the Junior Suite is worth it for a long stay and book it.","Booking for the Junior Suite is completed."
+hotel,The Last Minute Panic (Superior),"It's late and you need a room for tonight. 1) Search for a room for 1 person. 2) You recognize the ""Superior Room"" brand. 3) Click it. 4) Quickly verify check-in times or details. 5) Don't overthink it—book the Superior Room as fast as possible.","Booking for the Superior Room is completed."

docs/index.html

@@ -17,8 +17,8 @@
   <meta property="og:site_name" content="PHANTOM Research">
   <meta property="og:title" content="PHANTOM: Pricing Heuristics Against Non-human Transaction Orchestration Mechanisms">
   <meta property="og:description" content="Developing pricing heuristics to protect e-commerce platforms from systematic exploitation by LLM agents in dynamic pricing environments through behavioral signature detection.">
-  <meta property="og:url" content="TODO">
-  <meta property="og:image" content="TODO">
+  <meta property="og:url" content="https://velocitatem.github.io/PHANTOM/">
+  <meta property="og:image" content="https://raw.githubusercontent.com/velocitatem/PHANTOM/main/docs/static/images/carousel1.jpg">
   <meta property="og:image:width" content="1200">
   <meta property="og:image:height" content="630">
   <meta property="og:image:alt" content="PHANTOM Research Preview">
@@ -30,24 +30,19 @@
 
   <!-- Twitter -->
   <meta name="twitter:card" content="summary_large_image">
-  <!-- TODO: Replace with your lab/institution Twitter handle -->
-  <meta name="twitter:site" content="@YOUR_TWITTER_HANDLE">
-  <!-- TODO: Replace with first author's Twitter handle -->
-  <meta name="twitter:creator" content="@AUTHOR_TWITTER_HANDLE">
-  <!-- TODO: Same as paper title above -->
-  <meta name="twitter:title" content="PAPER_TITLE">
-  <!-- TODO: Same as description above -->
-  <meta name="twitter:description" content="BRIEF_DESCRIPTION_OF_YOUR_RESEARCH_CONTRIBUTION_AND_FINDINGS">
-  <!-- TODO: Same as social preview image above -->
-  <meta name="twitter:image" content="https://YOUR_DOMAIN.com/static/images/social_preview.png">
-  <meta name="twitter:image:alt" content="PAPER_TITLE - Research Preview">
+  <meta name="twitter:site" content="@velocitatem">
+  <meta name="twitter:creator" content="@velocitatem">
+  <meta name="twitter:title" content="PHANTOM: Pricing Heuristics Against Non-human Transaction Orchestration Mechanisms">
+  <meta name="twitter:description" content="A thesis project on defending dynamic pricing against LLM-driven reconnaissance and transaction orchestration.">
+  <meta name="twitter:image" content="https://raw.githubusercontent.com/velocitatem/PHANTOM/main/docs/static/images/carousel1.jpg">
+  <meta name="twitter:image:alt" content="PHANTOM research visual">
 
   <!-- Academic/Research Specific -->
   <meta name="citation_title" content="Pricing Heuristics Against Non-human Transaction Orchestration Mechanisms">
   <meta name="citation_author" content="Rösel, Daniel">
   <meta name="citation_publication_date" content="2025">
   <meta name="citation_conference_title" content="IE University Bachelor's Thesis">
-  <meta name="citation_pdf_url" content="TODO">
+  <meta name="citation_pdf_url" content="https://pub-d5b94a3c29fd40c6b3881946e463fdb7.r2.dev/thesis-latest.pdf">
 
   <!-- Additional SEO -->
   <meta name="theme-color" content="#2563eb">
@@ -103,50 +98,42 @@
   {
     "@context": "https://schema.org",
     "@type": "ScholarlyArticle",
-    "headline": "PAPER_TITLE",
-    "description": "BRIEF_DESCRIPTION_OF_YOUR_RESEARCH_CONTRIBUTION_AND_FINDINGS",
+    "headline": "PHANTOM: Pricing Heuristics Against Non-human Transaction Orchestration Mechanisms",
+    "description": "Research on preserving dynamic pricing integrity under LLM-mediated reconnaissance and purchasing behavior.",
     "author": [
       {
         "@type": "Person",
-        "name": "FIRST_AUTHOR_NAME",
+        "name": "Daniel Rösel",
         "affiliation": {
           "@type": "Organization",
-          "name": "INSTITUTION_NAME"
-        }
-      },
-      {
-        "@type": "Person",
-        "name": "SECOND_AUTHOR_NAME",
-        "affiliation": {
-          "@type": "Organization",
-          "name": "INSTITUTION_NAME"
+          "name": "IE University"
         }
       }
     ],
-    "datePublished": "2024-01-01",
+    "datePublished": "2025-01-01",
     "publisher": {
       "@type": "Organization",
-      "name": "CONFERENCE_OR_JOURNAL_NAME"
+      "name": "IE University"
     },
-    "url": "https://YOUR_DOMAIN.com/YOUR_PROJECT_PAGE",
-    "image": "https://YOUR_DOMAIN.com/static/images/social_preview.png",
-    "keywords": ["KEYWORD1", "KEYWORD2", "KEYWORD3", "machine learning", "computer vision"],
-    "abstract": "FULL_ABSTRACT_TEXT_HERE",
-    "citation": "BIBTEX_CITATION_HERE",
+    "url": "https://velocitatem.github.io/PHANTOM/",
+    "image": "https://raw.githubusercontent.com/velocitatem/PHANTOM/main/docs/static/images/carousel1.jpg",
+    "keywords": ["dynamic pricing", "llm agents", "e-commerce", "distributionally robust optimization", "reinforcement learning"],
+    "abstract": "This thesis formalizes Cost of Information erosion under agentic reconnaissance, learns separable human and agent behavior kernels, and trains contamination-aware robust pricing policies.",
+    "citation": "Rösel, Daniel. PHANTOM: Pricing Heuristics Against Non-human Transaction Orchestration Mechanisms. IE University, 2025.",
     "isAccessibleForFree": true,
     "license": "https://creativecommons.org/licenses/by/4.0/",
     "mainEntity": {
       "@type": "WebPage",
-      "@id": "https://YOUR_DOMAIN.com/YOUR_PROJECT_PAGE"
+      "@id": "https://velocitatem.github.io/PHANTOM/"
     },
     "about": [
       {
         "@type": "Thing",
-        "name": "RESEARCH_AREA_1"
+        "name": "Dynamic Pricing"
       },
       {
         "@type": "Thing",
-        "name": "RESEARCH_AREA_2"
+        "name": "Agent Behavior Modeling"
       }
     ]
   }
@@ -158,8 +145,7 @@
     "@context": "https://schema.org",
     "@type": "Organization",
     "name": "IE University",
-    "url": "https://www.ie.edu",
-    "logo": "TODO"
+    "url": "https://www.ie.edu"
   }
   </script>
 </head>
@@ -173,45 +159,72 @@
 
   <!-- More Works Dropdown -->
   <div class="more-works-container">
-    <button class="more-works-btn" onclick="toggleMoreWorks()" title="View More Works from Our Lab">
+    <button class="more-works-btn" onclick="toggleMoreWorks()" title="View project links and artifacts">
       <i class="fas fa-flask"></i>
-      More Works
+      Project Links
       <i class="fas fa-chevron-down dropdown-arrow"></i>
     </button>
     <div class="more-works-dropdown" id="moreWorksDropdown">
       <div class="dropdown-header">
-        <h4>More Works from Our Lab</h4>
+        <h4>Project Links</h4>
         <button class="close-btn" onclick="toggleMoreWorks()">
           <i class="fas fa-times"></i>
         </button>
       </div>
       <div class="works-list">
-        <!-- TODO: Replace with your lab's related works -->
-        <a href="https://arxiv.org/abs/PAPER_ID_1" class="work-item" target="_blank">
+        <a href="https://pub-d5b94a3c29fd40c6b3881946e463fdb7.r2.dev/thesis-latest.pdf" class="work-item" target="_blank">
           <div class="work-info">
-            <!-- TODO: Replace with actual paper title -->
-            <h5>Paper Title 1</h5>
-            <!-- TODO: Replace with brief description -->
-            <p>Brief description of the work and its main contribution.</p>
-            <!-- TODO: Replace with venue and year -->
-            <span class="work-venue">Conference/Journal 2024</span>
+            <h5>Thesis PDF</h5>
+            <p>Latest public build of the full thesis document.</p>
+            <span class="work-venue">IE University, 2025</span>
           </div>
           <i class="fas fa-external-link-alt"></i>
         </a>
-        <!-- TODO: Add more related works or remove extra items -->
-        <a href="https://arxiv.org/abs/PAPER_ID_2" class="work-item" target="_blank">
+        <a href="https://github.com/velocitatem/PHANTOM" class="work-item" target="_blank">
           <div class="work-info">
-            <h5>Paper Title 2</h5>
-            <p>Brief description of the work and its main contribution.</p>
-            <span class="work-venue">Conference/Journal 2023</span>
+            <h5>PHANTOM Repository</h5>
+            <p>Monorepo with paper source, platform code, and experiments.</p>
+            <span class="work-venue">Open Source</span>
           </div>
           <i class="fas fa-external-link-alt"></i>
         </a>
-        <a href="https://arxiv.org/abs/PAPER_ID_3" class="work-item" target="_blank">
+        <a href="https://github.com/velocitatem/p4p" class="work-item" target="_blank">
           <div class="work-info">
-            <h5>Paper Title 3</h5>
-            <p>Brief description of the work and its main contribution.</p>
-            <span class="work-venue">Conference/Journal 2023</span>
+            <h5>P4P Interaction Layer</h5>
+            <p>Reusable storefront and logging layer released for replication.</p>
+            <span class="work-venue">Public Artifact</span>
+          </div>
+          <i class="fas fa-external-link-alt"></i>
+        </a>
+        <a href="https://phantom-hotel.vercel.app" class="work-item" target="_blank">
+          <div class="work-info">
+            <h5>Hotel Mode Demo</h5>
+            <p>Public deployment of the hotel-style experiment interface.</p>
+            <span class="work-venue">Live Demo</span>
+          </div>
+          <i class="fas fa-external-link-alt"></i>
+        </a>
+        <a href="https://phantom-airline.vercel.app" class="work-item" target="_blank">
+          <div class="work-info">
+            <h5>Airline Mode Demo</h5>
+            <p>Public deployment of the airline-style experiment interface.</p>
+            <span class="work-venue">Live Demo</span>
+          </div>
+          <i class="fas fa-external-link-alt"></i>
+        </a>
+        <a href="https://blog.alves.world/series/phantom" class="work-item" target="_blank">
+          <div class="work-info">
+            <h5>Blog Series</h5>
+            <p>Behind-the-scenes posts covering thesis process, tooling, and insights.</p>
+            <span class="work-venue">To Boldly Code</span>
+          </div>
+          <i class="fas fa-external-link-alt"></i>
+        </a>
+        <a href="goals/README.md" class="work-item" target="_blank">
+          <div class="work-info">
+            <h5>Goal Library</h5>
+            <p>Task definitions used to assign actor objectives in experiments.</p>
+            <span class="work-venue">Experiment Design</span>
           </div>
           <i class="fas fa-external-link-alt"></i>
         </a>
@@ -233,14 +246,23 @@
 
                   <div class="is-size-5 publication-authors">
                     <span class="author-block">IE University<br>Bachelor's Thesis 2025</span>
-                    <span class="eql-cntrb"><small><br>Advisor: <a href="SECOND AUTHOR PERSONAL LINK" target="_blank">Alberto Martín Izquierdo</a></small></span>
+                    <span class="eql-cntrb"><small><br>Advisor: Alberto Martín Izquierdo</small></span>
                   </div>
 
                   <div class="column has-text-centered">
                     <div class="publication-links">
-                         <!-- TODO: Update with your arXiv paper ID -->
                       <span class="link-block">
-                        <a href="https://arxiv.org/pdf/<ARXIV PAPER ID>.pdf" target="_blank"
+                        <a href="https://blog.alves.world/series/phantom" target="_blank"
+                        class="external-link button is-normal is-rounded is-dark">
+                        <span class="icon">
+                          <i class="fas fa-blog"></i>
+                        </span>
+                        <span>Blog Series</span>
+                      </a>
+                    </span>
+
+                    <span class="link-block">
+                        <a href="https://pub-d5b94a3c29fd40c6b3881946e463fdb7.r2.dev/thesis-latest.pdf" target="_blank"
                         class="external-link button is-normal is-rounded is-dark">
                         <span class="icon">
                           <i class="fas fa-file-pdf"></i>
@@ -249,14 +271,13 @@
                       </a>
                     </span>
 
-                    <!-- TODO: Add your supplementary material PDF or remove this section -->
                     <span class="link-block">
-                      <a href="static/pdfs/supplementary_material.pdf" target="_blank"
+                      <a href="https://huggingface.co/datasets/velocitatem/whoclickedit" target="_blank"
                       class="external-link button is-normal is-rounded is-dark">
                       <span class="icon">
-                        <i class="fas fa-file-pdf"></i>
+                        <i class="fas fa-database"></i>
                       </span>
-                      <span>Supplementary</span>
+                      <span>Dataset</span>
                     </a>
                   </span>
 
@@ -270,42 +291,43 @@
                   </a>
                 </span>
 
-                <!-- TODO: Update with your arXiv paper ID -->
                 <span class="link-block">
-                  <a href="https://arxiv.org/abs/<ARXIV PAPER ID>" target="_blank"
+                  <a href="https://phantom-hotel.vercel.app" target="_blank"
                   class="external-link button is-normal is-rounded is-dark">
                   <span class="icon">
-                    <i class="ai ai-arxiv"></i>
+                    <i class="fas fa-globe"></i>
                   </span>
-                  <span>arXiv</span>
+                  <span>Hotel Demo</span>
                 </a>
               </span>
+
+              <span class="link-block">
+                <a href="https://phantom-airline.vercel.app" target="_blank"
+                class="external-link button is-normal is-rounded is-dark">
+                <span class="icon">
+                  <i class="fas fa-plane"></i>
+                </span>
+                <span>Airline Demo</span>
+              </a>
+            </span>
             </div>
           </div>
         </div>
       </div>
     </div>
-  </div>
 </section>
 
 
-<!-- Teaser video-->
 <section class="hero teaser">
   <div class="container is-max-desktop">
     <div class="hero-body">
-      <!-- TODO: Replace with your teaser video -->
-      <video poster="" id="tree" autoplay controls muted loop height="100%" preload="metadata">
-        <!-- TODO: Add your video file path here -->
-        <source src="static/videos/banner_video.mp4" type="video/mp4">
-      </video>
-      <!-- TODO: Replace with your video description -->
-      <h2 class="subtitle has-text-centered">
-        Aliquam vitae elit ullamcorper tellus egestas pellentesque. Ut lacus tellus, maximus vel lectus at, placerat pretium mi. Maecenas dignissim tincidunt vestibulum. Sed consequat hendrerit nisl ut maximus.
-      </h2>
+      <div class="publication-banner">
+        <img src="static/images/banner.svg" alt="PHANTOM teaser diagram connecting vulnerability, behavioral signal, and robust control" width="1920" height="1080" decoding="async" style="display:block; width:100%; height:auto;" onerror="this.onerror=null;this.src='static/images/carousel2.jpg';"/>
+      </div>
     </div>
   </div>
 </section>
-<!-- End teaser video -->
+
 
 <!-- Paper abstract -->
 <section class="section hero is-light">
@@ -315,7 +337,10 @@
         <h2 class="title is-3">Abstract</h2>
         <div class="content has-text-justified">
           <p>
-            The primary objective of this thesis is to develop and validate pricing heuristics that protect e-commerce platforms from systematic exploitation by Large Language Model (LLM) agents within dynamic pricing environments. As AI agents increasingly mediate consumer transactions, they enable users to circumvent the Cost of Information (the price premium accumulated through demand signal expression) by conducting reconnaissance in isolated sessions before executing purchases through clean sessions at base prices. This research will make an anticipatory contribution by adapting recommendation system methodologies to distinguish between genuine human browsing behaviour and agent-orchestrated information gathering, thereby enabling pricing systems to maintain margin integrity without degrading the user experience for legitimate customers or getting rid of leads generated by LLMs.
+            When you shop online, prices often change based on how much interest you show — the more you browse, the more the site learns about your intent and may raise prices accordingly. This works because stores assume that a curious, engaged shopper is more likely to buy. But AI assistants are now doing the shopping research on behalf of users: they browse in one session to gather price information and then let the user purchase in a fresh session at the lower, unadjusted price. The store never sees the connection between the two, so it never gets to factor in that genuine intent — and loses the revenue it would have earned.
+          </p>
+          <p>
+            PHANTOM studies this problem and builds defenses against it. We created a realistic fake store (in hotel and airline modes) where both real people and AI agents were given shopping tasks, and we recorded every click, scroll, and page visit. By comparing how humans and AI agents move through a site, we found clear patterns that tell them apart. We then used those patterns to build a smarter pricing system that can recognize when it is likely talking to an AI scout and adjust its strategy accordingly — protecting the store's margins without making things worse for genuine shoppers.
           </p>
         </div>
       </div>
@@ -324,97 +349,90 @@
 </section>
 <!-- End paper abstract -->
 
+<section class="section">
+  <div class="container is-max-desktop">
+    <div class="content has-text-justified">
+      <h2 class="title is-3 has-text-centered">Project Scope</h2>
+      <p>
+        The current thesis revision extends both theory and implementation. The main research question is how a pricing system can preserve margin integrity when browsing and purchasing are increasingly orchestrated by AI agents.
+      </p>
+      <ul>
+        <li>Formal contribution: a Cost of Information erosion theorem showing why price-query saturation can collapse dynamic pricing power.</li>
+        <li>System contribution: a hybrid online/offline stack (Next.js storefront, pricing provider, Kafka event streams, Airflow ETL, Redis serving layer).</li>
+        <li>Modeling contribution: class-specific transition kernels for human and agent behavior, with KL-divergence based separability scores.</li>
+        <li>Control contribution: a contamination-aware DR-RL pricing policy trained under distributional uncertainty using Wasserstein-style robustness.</li>
+      </ul>
+      <p>
+        Controlled trials currently include balanced human and agent sessions with goal-driven tasks across hotel and airline interfaces. Early separability results are strong (Mann-Whitney U=2.0, p=0.0006), while robust pricing gains remain regime-dependent and are being calibrated in larger sweeps.
+      </p>
+    </div>
+  </div>
+</section>
+
 
 <!-- Image carousel -->
+<!--
 <section class="hero is-small">
   <div class="hero-body">
     <div class="container">
       <div id="results-carousel" class="carousel results-carousel">
        <div class="item">
-        <!-- TODO: Replace with your research result images -->
         <img src="static/images/carousel1.jpg" alt="First research result visualization" loading="lazy"/>
-        <!-- TODO: Replace with description of this result -->
         <h2 class="subtitle has-text-centered">
-          First image description.
+          Early simulator traces showing how policy choice can push prices toward aggressive high-end regimes.
         </h2>
       </div>
       <div class="item">
-        <!-- Your image here -->
         <img src="static/images/carousel2.jpg" alt="Second research result visualization" loading="lazy"/>
         <h2 class="subtitle has-text-centered">
-          Second image description.
+          Human and agent behavior diverge at the transition-kernel level, enabling usable session-level separability.
         </h2>
       </div>
       <div class="item">
-        <!-- Your image here -->
         <img src="static/images/carousel3.jpg" alt="Third research result visualization" loading="lazy"/>
         <h2 class="subtitle has-text-centered">
-         Third image description.
+         End-to-end architecture linking web interactions, pricing queries, event streams, and model updates.
        </h2>
      </div>
      <div class="item">
-      <!-- Your image here -->
       <img src="static/images/carousel4.jpg" alt="Fourth research result visualization" loading="lazy"/>
       <h2 class="subtitle has-text-centered">
-        Fourth image description.
+        Contamination-aware evaluation compares robust and non-robust pricing behavior across alpha sweeps.
       </h2>
     </div>
   </div>
 </div>
 </div>
 </section>
+-->
 <!-- End image carousel -->
 
 
 
 
-<!-- Youtube video -->
-<section class="hero is-small is-light">
-  <div class="hero-body">
-    <div class="container">
-      <!-- Paper video. -->
-      <h2 class="title is-3">Video Presentation</h2>
-      <div class="columns is-centered has-text-centered">
-        <div class="column is-four-fifths">
-
-          <div class="publication-video">
-            <!-- TODO: Replace with your YouTube video ID -->
-            <iframe src="https://www.youtube.com/embed/JkaxUblCGz0" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>
-          </div>
-        </div>
-      </div>
-    </div>
-  </div>
-</section>
-<!-- End youtube video -->
-
-
 <!-- Video carousel -->
 <section class="hero is-small">
   <div class="hero-body">
     <div class="container">
-      <h2 class="title is-3">Another Carousel</h2>
-      <div id="results-carousel" class="carousel results-carousel">
+      <h2 class="title is-3">Defense Scenes</h2>
+      <div id="videos-carousel" class="carousel results-carousel">
         <div class="item item-video1">
-          <!-- TODO: Add poster image for better preview -->
           <video poster="" id="video1" controls muted loop height="100%" preload="metadata">
-            <!-- Your video file here -->
-            <source src="static/videos/carousel1.mp4" type="video/mp4">
+            <source src="static/videos/COIFirstPrinciplesScene.mp4" type="video/mp4">
           </video>
+          <h2 class="subtitle has-text-centered">COI from first principles.</h2>
         </div>
         <div class="item item-video2">
-          <!-- TODO: Add poster image for better preview -->
           <video poster="" id="video2" controls muted loop height="100%" preload="metadata">
-            <!-- Your video file here -->
-            <source src="static/videos/carousel2.mp4" type="video/mp4">
+            <source src="static/videos/BehaviorKernelConstructionScene.mp4" type="video/mp4">
           </video>
+          <h2 class="subtitle has-text-centered">Behavioral kernel construction: learning how humans and agents differ.</h2>
         </div>
         <div class="item item-video3">
-          <!-- TODO: Add poster image for better preview -->
           <video poster="" id="video3" controls muted loop height="100%" preload="metadata">
-            <!-- Your video file here -->
-            <source src="static/videos/carousel3.mp4" type="video/mp4">
+            <source src="static/videos/RobustControlScene.mp4" type="video/mp4">
           </video>
+          <h2 class="subtitle has-text-centered">Distributionally robust control loop.</h2>
         </div>
       </div>
     </div>
@@ -430,11 +448,10 @@
 <!-- Paper poster -->
 <section class="hero is-small is-light">
   <div class="hero-body">
-    <div class="container">
-      <h2 class="title">Poster</h2>
+      <div class="container">
+      <h2 class="title">Full Thesis</h2>
 
-      <!-- TODO: Replace with your poster PDF -->
-      <iframe  src="static/pdfs/sample.pdf" width="100%" height="550">
+      <iframe title="PHANTOM thesis PDF" src="https://pub-d5b94a3c29fd40c6b3881946e463fdb7.r2.dev/thesis-latest.pdf" width="100%" height="550">
           </iframe>
 
       </div>
@@ -456,7 +473,7 @@
       </div>
       <pre id="bibtex-code"><code>@thesis{Rosel2025PHANTOM,
   title={Pricing Heuristics Against Non-human Transaction Orchestration Mechanisms},
-  author={R{\"o}sel, Daniel},
+  author={Rösel, Daniel},
   school={IE University},
   year={2025},
   address={Madrid, Spain},

docs/static/images/banner.svg

@@ -0,0 +1,246 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1920 1080" width="1920" height="1080" style="background-color: #FAFAFA; font-family: 'Helvetica Neue', Helvetica, Arial, sans-serif;">
+    <defs>
+        <!-- Soft Drop Shadow for Panels -->
+        <filter id="shadow" x="-10%" y="-10%" width="130%" height="130%">
+            <feDropShadow dx="2" dy="4" stdDeviation="6" flood-color="#000000" flood-opacity="0.06"/>
+        </filter>
+        <filter id="light-shadow" x="-5%" y="-5%" width="110%" height="110%">
+            <feDropShadow dx="1" dy="2" stdDeviation="2" flood-color="#000000" flood-opacity="0.04"/>
+        </filter>
+
+        <!-- Arrowhead Marker -->
+        <marker id="arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="6" markerHeight="6" orient="auto-start-reverse">
+            <path d="M 0 0 L 10 5 L 0 10 z" fill="#888888" />
+        </marker>
+        <marker id="arrow-dark" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="6" markerHeight="6" orient="auto-start-reverse">
+            <path d="M 0 0 L 10 5 L 0 10 z" fill="#555555" />
+        </marker>
+    </defs>
+
+    <!-- COLUMN DIVIDERS -->
+    <line x1="640" y1="60" x2="640" y2="1020" stroke="#EAEAEA" stroke-width="2" stroke-dasharray="10,10"/>
+    <line x1="1280" y1="60" x2="1280" y2="1020" stroke="#EAEAEA" stroke-width="2" stroke-dasharray="10,10"/>
+
+
+    <!-- ========================================================= -->
+    <!-- COLUMN 1: THE THREAT (COI & SATURATION)                   -->
+    <!-- ========================================================= -->
+    <text x="60" y="80" font-family="Georgia, serif" font-size="28" font-weight="bold" fill="#333333">1. The Vulnerability</text>
+    <line x1="60" y1="100" x2="580" y2="100" stroke="#DDDDDD" stroke-width="2"/>
+
+    <!-- Top: COI Bell Curve -->
+    <g transform="translate(60, 130)">
+        <text x="0" y="30" font-size="24" font-weight="bold" fill="#444">Cost of Information from First Principles</text>
+        <text x="0" y="70" font-family="Georgia, serif" font-style="italic" font-size="22" fill="#8C7A6B">P ~ π(τ)</text>
+        <text x="0" y="105" font-family="Georgia, serif" font-style="italic" font-size="22" fill="#8C7A6B"><tspan text-decoration="underline">p</tspan> = reservation price</text>
+        <text x="0" y="140" font-family="Georgia, serif" font-style="italic" font-size="22" fill="#8C7A6B">M = P - <tspan text-decoration="underline">p</tspan></text>
+
+        <!-- Bell Curve -->
+        <path d="M 40 340 C 140 340, 160 160, 260 160 C 360 160, 380 340, 480 340" stroke="#3AB09E" stroke-width="5" fill="none"/>
+        <line x1="40" y1="340" x2="500" y2="340" stroke="#333" stroke-width="2"/>
+
+        <!-- Markers p and E[P] -->
+        <line x1="150" y1="340" x2="150" y2="160" stroke="#E37862" stroke-width="2" stroke-dasharray="6,4"/>
+        <text x="150" y="375" font-family="Georgia" font-style="italic" font-size="22" fill="#E37862" text-anchor="middle">p</text>
+
+        <line x1="260" y1="340" x2="260" y2="160" stroke="#85B589" stroke-width="2" stroke-dasharray="6,4"/>
+        <text x="260" y="375" font-family="Georgia" font-style="italic" font-size="22" fill="#85B589" text-anchor="middle">E[P]</text>
+
+        <!-- COI Annotation -->
+        <line x1="150" y1="150" x2="260" y2="150" stroke="#E37862" stroke-width="2" marker-start="url(#arrow)" marker-end="url(#arrow)"/>
+        <text x="310" y="138" font-size="16" fill="#E37862" text-anchor="middle">average information rent</text>
+        <text x="310" y="118" font-family="Georgia" font-style="italic" font-size="22" fill="#E37862" font-weight="bold" text-anchor="middle">COI := E[P] - p</text>
+    </g>
+
+    <!-- Bottom: Agent Saturation -->
+    <g transform="translate(60, 580)">
+        <text x="0" y="30" font-size="24" font-weight="bold" fill="#444">Why COI Erodes with Agent Saturation</text>
+        <text x="0" y="75" font-family="Georgia, serif" font-style="italic" font-size="22" fill="#8C7A6B">p<tspan font-size="14" dy="5">(1)</tspan><tspan dy="-5"> = min(p</tspan><tspan font-size="14" dy="5">1</tspan><tspan dy="-5">, ..., p</tspan><tspan font-size="14" dy="5">N</tspan><tspan dy="-5">)</tspan></text>
+        <text x="0" y="115" font-family="Georgia, serif" font-style="italic" font-size="22" fill="#8C7A6B">P(p<tspan font-size="14" dy="5">(1)</tspan><tspan dy="-5"> > t) = [1 - F(t)]</tspan><tspan font-size="14" dy="-10">N</tspan></text>
+
+        <!-- Erosion Graph -->
+        <rect x="120" y="150" width="280" height="230" fill="#FFFFFF" filter="url(#shadow)" rx="8"/>
+        <line x1="140" y1="350" x2="380" y2="350" stroke="#333" stroke-width="2"/>
+        <line x1="140" y1="350" x2="140" y2="170" stroke="#333" stroke-width="2"/>
+        <text x="260" y="375" font-size="16" font-style="italic" fill="#555" text-anchor="middle">F(t)</text>
+        <text x="120" y="260" font-size="16" font-style="italic" fill="#555" text-anchor="middle" transform="rotate(-90 120 260)">[1 - F(t)]^N</text>
+
+        <!-- Curves -->
+        <path d="M 140 170 C 220 250, 300 320, 380 350" stroke="#4EA5D9" stroke-width="3" fill="none"/>
+        <text x="390" y="220" font-size="16" fill="#4EA5D9" font-weight="bold">N=1</text>
+
+        <path d="M 140 170 C 180 260, 240 330, 380 350" stroke="#85B589" stroke-width="3" fill="none"/>
+        <text x="390" y="250" font-size="16" fill="#85B589" font-weight="bold">N=4</text>
+
+        <path d="M 140 170 C 150 290, 180 340, 380 350" stroke="#E37862" stroke-width="3" fill="none"/>
+        <text x="390" y="280" font-size="16" fill="#E37862" font-weight="bold">N=16</text>
+
+        <text x="260" y="420" font-size="20" fill="#555" text-anchor="middle">As independent query count grows,</text>
+        <text x="260" y="445" font-size="20" fill="#E37862" font-weight="bold" text-anchor="middle">realizable markup collapses.</text>
+    </g>
+
+
+    <!-- ========================================================= -->
+    <!-- COLUMN 2: THE BEHAVIORAL SIGNAL                           -->
+    <!-- ========================================================= -->
+    <text x="700" y="80" font-family="Georgia, serif" font-size="28" font-weight="bold" fill="#333333">2. The Behavioral Signals</text>
+    <line x1="700" y1="100" x2="1220" y2="100" stroke="#DDDDDD" stroke-width="2"/>
+
+    <!-- Top: Transition Kernels -->
+    <g transform="translate(700, 130)">
+        <text x="0" y="30" font-size="24" font-weight="bold" fill="#444">From Session Paths to Transition Kernels</text>
+
+        <text x="0" y="75" font-size="20" fill="#85B589" font-weight="bold">human: start → view → detail → cart → purchase</text>
+        <text x="0" y="115" font-size="20" fill="#E37862" font-weight="bold">agent: start → view → detail → view → detail</text>
+
+        <text x="0" y="170" font-family="Georgia, serif" font-style="italic" font-size="24" fill="#8C7A6B">
+            P&#770;(s'|s) = <tspan font-size="18" dy="-12">N(s,s')</tspan> / <tspan font-size="18" dy="12">Σ N(s,k)</tspan>
+        </text>
+
+        <!-- Matrix Representation -->
+        <rect x="0" y="220" width="500" height="180" fill="#FFFFFF" filter="url(#shadow)" rx="8"/>
+
+        <text x="125" y="250" font-size="16" fill="#4EA5D9" text-anchor="middle">transition counts N(s,s')</text>
+        <text x="375" y="250" font-size="16" fill="#85B589" text-anchor="middle">normalized kernel T</text>
+
+        <!-- Matrix 1 -->
+        <g transform="translate(45, 270)">
+            <rect x="-6" y="-8" width="172" height="128" rx="6" fill="none" stroke="#DDDDDD" stroke-width="1.5"/>
+            <path d="M 10 0 L 0 0 L 0 110 L 10 110 M 150 0 L 160 0 L 160 110 L 150 110" stroke="#A0A0A0" stroke-width="2.5" fill="none"/>
+            <text x="80" y="20" font-family="monospace" font-size="14" fill="#555" text-anchor="middle" textLength="142" lengthAdjust="spacingAndGlyphs">0.00 8.00 0.00 0.00</text>
+            <text x="80" y="50" font-family="monospace" font-size="14" fill="#555" text-anchor="middle" textLength="142" lengthAdjust="spacingAndGlyphs">0.00 2.00 5.00 1.00</text>
+            <text x="80" y="80" font-family="monospace" font-size="14" fill="#555" text-anchor="middle" textLength="142" lengthAdjust="spacingAndGlyphs">0.00 3.00 2.00 4.00</text>
+            <text x="80" y="110" font-family="monospace" font-size="14" fill="#555" text-anchor="middle" textLength="142" lengthAdjust="spacingAndGlyphs">0.00 1.00 0.00 6.00</text>
+        </g>
+
+        <!-- Arrow -->
+        <line x1="225" y1="320" x2="265" y2="320" stroke="#999" stroke-width="3" marker-end="url(#arrow-dark)"/>
+
+        <!-- Matrix 2 -->
+        <g transform="translate(295, 270)">
+            <rect x="-6" y="-8" width="172" height="128" rx="6" fill="none" stroke="#DDDDDD" stroke-width="1.5"/>
+            <path d="M 10 0 L 0 0 L 0 110 L 10 110 M 150 0 L 160 0 L 160 110 L 150 110" stroke="#A0A0A0" stroke-width="2.5" fill="none"/>
+            <text x="80" y="20" font-family="monospace" font-size="14" fill="#333" text-anchor="middle" textLength="142" lengthAdjust="spacingAndGlyphs">0.00 1.00 0.00 0.00</text>
+            <text x="80" y="50" font-family="monospace" font-size="14" fill="#333" text-anchor="middle" textLength="142" lengthAdjust="spacingAndGlyphs">0.00 0.25 0.62 0.13</text>
+            <text x="80" y="80" font-family="monospace" font-size="14" fill="#333" text-anchor="middle" textLength="142" lengthAdjust="spacingAndGlyphs">0.00 0.33 0.22 0.45</text>
+            <text x="80" y="110" font-family="monospace" font-size="14" fill="#333" text-anchor="middle" textLength="142" lengthAdjust="spacingAndGlyphs">0.00 0.14 0.00 0.86</text>
+        </g>
+
+        <text x="250" y="440" font-size="18" fill="#777" text-anchor="middle">Kernel shape is the compact behavioral signature used downstream.</text>
+    </g>
+
+    <!-- Bottom: Separability Distributions -->
+    <g transform="translate(700, 600)">
+        <text x="0" y="30" font-size="24" font-weight="bold" fill="#444">Separability into a Control Signal</text>
+        <text x="0" y="75" font-family="Georgia, serif" font-style="italic" font-size="22" fill="#8C7A6B">Δ<tspan font-size="16" dy="5">H</tspan><tspan dy="-5"> = D</tspan><tspan font-size="16" dy="5">KL</tspan><tspan dy="-5">(T&#770;' || T&#772;</tspan><tspan font-size="16" dy="5">H</tspan><tspan dy="-5">)</tspan></text>
+        <text x="0" y="115" font-family="Georgia, serif" font-style="italic" font-size="22" fill="#8C7A6B">Δ<tspan font-size="16" dy="5">A</tspan><tspan dy="-5"> = D</tspan><tspan font-size="16" dy="5">KL</tspan><tspan dy="-5">(T&#770;' || T&#772;</tspan><tspan font-size="16" dy="5">A</tspan><tspan dy="-5">)</tspan></text>
+        <text x="0" y="155" font-family="Georgia, serif" font-style="italic" font-size="24" fill="#8C7A6B">g = Δ<tspan font-size="16" dy="5">H</tspan><tspan dy="-5"> - Δ</tspan><tspan font-size="16" dy="5">A</tspan></text>
+
+        <!-- Curves -->
+        <g transform="translate(80, 160)">
+            <line x1="0" y1="200" x2="360" y2="200" stroke="#333" stroke-width="2"/>
+            <text x="180" y="235" font-family="Georgia, serif" font-style="italic" font-size="22" text-anchor="middle">g = Δ<tspan font-size="16" dy="5">H</tspan><tspan dy="-5"> - Δ</tspan><tspan font-size="16" dy="5">A</tspan></text>
+
+            <!-- Human Curve -->
+            <path d="M 0 200 C 50 200, 80 40, 130 40 C 180 40, 210 200, 260 200" stroke="#4EA5D9" stroke-width="5" fill="none"/>
+            <text x="70" y="110" font-size="22" fill="#4EA5D9" font-weight="bold">human</text>
+
+            <!-- Agent Curve -->
+            <path d="M 100 200 C 150 200, 180 40, 230 40 C 280 40, 310 200, 360 200" stroke="#E37862" stroke-width="5" fill="none"/>
+            <text x="290" y="110" font-size="22" fill="#E37862" font-weight="bold">agent</text>
+
+            <!-- Decision Boundary -->
+            <line x1="180" y1="200" x2="180" y2="10" stroke="#999" stroke-width="2" stroke-dasharray="8,5"/>
+            <text x="180" y="-5" font-size="16" fill="#777" text-anchor="middle">decision boundary</text>
+
+            <circle cx="210" cy="200" r="6" fill="#ECA233"/>
+            <text x="210" y="180" font-family="Georgia" font-style="italic" font-size="20" fill="#ECA233" text-anchor="middle">g_obs</text>
+
+            <text x="180" y="280" font-size="18" fill="#555" text-anchor="middle">Positive gap shifts score toward agent traffic.</text>
+        </g>
+    </g>
+
+
+    <!-- ========================================================= -->
+    <!-- COLUMN 3: THE SOLUTION (CONTAMINATION & DR-RL)            -->
+    <!-- ========================================================= -->
+    <text x="1340" y="80" font-family="Georgia, serif" font-size="28" font-weight="bold" fill="#333333">3. Robust Control &amp; Contamination</text>
+    <line x1="1340" y1="100" x2="1860" y2="100" stroke="#DDDDDD" stroke-width="2"/>
+
+    <!-- Top: Contamination Generator -->
+    <g transform="translate(1340, 130)">
+        <text x="0" y="30" font-size="24" font-weight="bold" fill="#444">Contamination Generator G(α)</text>
+
+        <!-- Boxes -->
+        <rect x="20" y="70" width="200" height="50" fill="#D0E5E0" filter="url(#shadow)" rx="6"/>
+        <text x="120" y="100" font-size="18" fill="#222" text-anchor="middle">labeled human sessions</text>
+
+        <rect x="280" y="70" width="200" height="50" fill="#EAD0C8" filter="url(#shadow)" rx="6"/>
+        <text x="380" y="100" font-size="18" fill="#222" text-anchor="middle">synthetic agent sessions</text>
+
+        <!-- Arrows -->
+        <line x1="120" y1="130" x2="200" y2="180" stroke="#888" stroke-width="3" marker-end="url(#arrow-dark)"/>
+        <line x1="380" y1="130" x2="300" y2="180" stroke="#888" stroke-width="3" marker-end="url(#arrow-dark)"/>
+
+        <!-- Mixed Batch -->
+        <rect x="150" y="190" width="200" height="50" fill="#F4E9CD" filter="url(#shadow)" rx="6"/>
+        <text x="250" y="220" font-size="18" fill="#222" text-anchor="middle">mixed batch for training</text>
+
+        <!-- Alpha Bar -->
+        <text x="250" y="275" font-family="Georgia, serif" font-size="20" fill="#555" text-anchor="middle">alpha = 0.33</text>
+
+        <rect x="50" y="290" width="268" height="30" fill="#4EA5D9"/>
+        <rect x="318" y="290" width="132" height="30" fill="#E37862"/>
+        <text x="184" y="340" font-size="18" fill="#4EA5D9" text-anchor="middle">human share (1-α)</text>
+        <text x="384" y="340" font-size="18" fill="#E37862" text-anchor="middle">agent share (α)</text>
+    </g>
+
+    <!-- Bottom: Distributionally Robust Control -->
+    <g transform="translate(1340, 600)">
+        <text x="0" y="30" font-size="24" font-weight="bold" fill="#444">Distributionally Robust Control Layer</text>
+        <text x="0" y="80" font-family="Georgia, serif" font-style="italic" font-size="22" fill="#8C7A6B">
+            π* = arg max<tspan font-size="16" dy="5">π</tspan> min<tspan font-size="16" dy="0">Q ∈ U<tspan font-size="12" dy="5">ε</tspan></tspan>
+            <tspan dy="-10"> E</tspan><tspan font-size="16" dy="5">d ~ Q</tspan>
+            <tspan dy="-5">[ R(p,d) - λ COI</tspan><tspan font-size="16" dy="5">leak</tspan><tspan dy="-5">(p,τ') ]</tspan>
+        </text>
+
+        <!-- Ambiguity Ball -->
+        <g transform="translate(140, 260)">
+            <line x1="-130" y1="0" x2="130" y2="0" stroke="#CCC" stroke-width="2"/>
+            <line x1="0" y1="-130" x2="0" y2="130" stroke="#CCC" stroke-width="2"/>
+
+            <circle cx="0" cy="0" r="110" stroke="#C4A45B" stroke-width="4" fill="rgba(196,164,91,0.06)"/>
+            <text x="-95" y="-120" font-family="Georgia" font-style="italic" font-size="24" fill="#C4A45B">U<tspan font-size="16" dy="5">ε</tspan></text>
+
+            <!-- Points -->
+            <circle cx="0" cy="0" r="7" fill="#4EA5D9"/>
+            <text x="12" y="24" font-family="Georgia" font-style="italic" font-size="22" fill="#4EA5D9">P&#770;<tspan font-size="14" dy="5">N</tspan></text>
+
+            <circle cx="-60" cy="-40" r="7" fill="#E37862"/>
+            <text x="-140" y="-50" font-family="Georgia" font-style="italic" font-size="18" fill="#E37862">worst-case Q*</text>
+
+            <circle cx="50" cy="-70" r="6" fill="#85B589"/>
+            <circle cx="70" cy="50" r="6" fill="#85B589"/>
+            <circle cx="-40" cy="80" r="6" fill="#85B589"/>
+        </g>
+
+        <!-- Process Steps -->
+        <g transform="translate(320, 140)">
+            <rect x="0" y="0" width="220" height="45" fill="#FDEFEF" filter="url(#light-shadow)" rx="6"/>
+            <text x="110" y="28" font-size="16" fill="#E37862" font-weight="bold" text-anchor="middle">inner min picks Q*</text>
+
+            <line x1="110" y1="55" x2="110" y2="85" stroke="#999" stroke-width="2" marker-end="url(#arrow-dark)"/>
+
+            <rect x="0" y="95" width="220" height="45" fill="#F4E9CD" filter="url(#light-shadow)" rx="6"/>
+            <text x="110" y="123" font-size="16" fill="#9E8033" font-weight="bold" text-anchor="middle">sample demand from Q*</text>
+
+            <line x1="110" y1="150" x2="110" y2="180" stroke="#999" stroke-width="2" marker-end="url(#arrow-dark)"/>
+
+            <rect x="0" y="190" width="220" height="45" fill="#E6F2ED" filter="url(#light-shadow)" rx="6"/>
+            <text x="110" y="218" font-size="16" fill="#428062" font-weight="bold" text-anchor="middle">outer max updates policy</text>
+        </g>
+
+        <text x="250" y="440" font-size="18" fill="#555" text-anchor="middle">Reward is evaluated on demand drawn from Q*, then used for the policy step.</text>
+    </g>
+
+</svg>

engine/backends/__init__.py

@@ -0,0 +1 @@
+__all__ = ["evaluate", "make_env", "train_qtable", "train_sb3"]

engine/backends/common.py

@@ -0,0 +1,181 @@
+from __future__ import annotations
+
+from typing import Any, Mapping
+
+import numpy as np
+
+
+def make_env(cfg: Mapping[str, Any]):
+    from gymnasium.wrappers import FlattenObservation
+
+    from ..lib.wrappers import EconomicMetricsWrapper
+    from ..wrapper import PHANTOM
+
+    env = PHANTOM(
+        n_products=int(cfg["n_products"]),
+        alpha=float(cfg["alpha"]),
+        N=int(cfg["N"]),
+        agent_params=(
+            float(cfg.get("agent_mu", 45.0)),
+            float(cfg.get("agent_std", 15.0)),
+        ),
+        price_bounds=(float(cfg["price_low"]), float(cfg["price_high"])),
+        lambda_coi=float(cfg["lambda_coi"]),
+        robust_radius=float(cfg["robust_radius"]),
+        robust_points=int(cfg["robust_points"]),
+        robust_rollouts=int(cfg.get("robust_rollouts", 1)),
+        info_value=float(cfg["info_value"]),
+        eta_ux=float(cfg.get("eta_ux", 0.5)),
+        reward_profit_weight=float(cfg.get("reward_profit_weight", 1.0)),
+        action_levels=int(cfg["action_levels"]),
+        action_scale_low=float(cfg["action_scale_low"]),
+        action_scale_high=float(cfg["action_scale_high"]),
+        max_steps=int(cfg.get("max_steps", 100)),
+        margin_floor=float(cfg.get("margin_floor", 0.05)),
+        margin_floor_patience=int(cfg.get("margin_floor_patience", 5)),
+        render_mode=None,
+    )
+    env = EconomicMetricsWrapper(env)
+    return FlattenObservation(env)
+
+
+def _action(agent: Any, obs: Any, deterministic: bool = True):
+    out = agent.predict(obs, deterministic=deterministic)
+    action = out[0] if isinstance(out, tuple) else out
+    if isinstance(action, np.ndarray) and action.size == 1:
+        return int(action.reshape(-1)[0])
+    return action
+
+
+def _evaluate_env(agent: Any, env: Any, episodes: int) -> dict[str, float]:
+    rewards: list[float] = []
+    revenues: list[float] = []
+    margins: list[float] = []
+    coi_levels: list[float] = []
+    coi_leakages: list[float] = []
+    volatilities: list[float] = []
+    upward_volatilities: list[float] = []
+    supra_shares: list[float] = []
+    supra_penalties: list[float] = []
+    agent_probs: list[float] = []
+
+    for _ in range(int(episodes)):
+        obs, _ = env.reset()
+        done = False
+        ep_reward = 0.0
+        ep_revenue = 0.0
+        ep_margin = 0.0
+        ep_coi = 0.0
+        ep_coi_leakage = 0.0
+        ep_volatility = 0.0
+        ep_upward_volatility = 0.0
+        ep_supra_share = 0.0
+        ep_supra_penalty = 0.0
+        ep_agent_prob = 0.0
+        steps = 0
+
+        while not done:
+            obs, reward, term, trunc, info = env.step(_action(agent, obs, True))
+            done = bool(term or trunc)
+            econ = info.get("economics", {})
+            ep_reward += float(reward)
+            ep_revenue += float(econ.get("revenue", info.get("revenue", 0.0)))
+            ep_margin += float(econ.get("margin", 0.0))
+            ep_coi += float(econ.get("coi_level", 0.0))
+            ep_coi_leakage += float(econ.get("coi_leakage", 0.0))
+            ep_volatility += float(econ.get("volatility", 0.0))
+            ep_upward_volatility += float(
+                info.get("upward_volatility", econ.get("upward_volatility", 0.0))
+            )
+            ep_supra_share += float(
+                info.get("supra_share", econ.get("supra_share", 0.0))
+            )
+            ep_supra_penalty += float(
+                info.get("supra_penalty", econ.get("supra_penalty", 0.0))
+            )
+            ep_agent_prob += float(econ.get("agent_prob", info.get("agent_prob", 0.0)))
+            steps += 1
+
+        rewards.append(ep_reward)
+        revenues.append(ep_revenue)
+        denom = max(steps, 1)
+        margins.append(ep_margin / denom)
+        coi_levels.append(ep_coi / denom)
+        coi_leakages.append(ep_coi_leakage / denom)
+        volatilities.append(ep_volatility / denom)
+        upward_volatilities.append(ep_upward_volatility / denom)
+        supra_shares.append(ep_supra_share / denom)
+        supra_penalties.append(ep_supra_penalty / denom)
+        agent_probs.append(ep_agent_prob / denom)
+
+    return {
+        "eval/reward_mean": float(np.mean(rewards)) if rewards else 0.0,
+        "eval/reward_std": float(np.std(rewards)) if rewards else 0.0,
+        "eval/revenue_mean": float(np.mean(revenues)) if revenues else 0.0,
+        "eval/revenue_std": float(np.std(revenues)) if revenues else 0.0,
+        "eval/margin_mean": float(np.mean(margins)) if margins else 0.0,
+        "eval/coi_level_mean": float(np.mean(coi_levels)) if coi_levels else 0.0,
+        "eval/coi_leakage_mean": float(np.mean(coi_leakages)) if coi_leakages else 0.0,
+        "eval/volatility_mean": float(np.mean(volatilities)) if volatilities else 0.0,
+        "eval/upward_volatility_mean": (
+            float(np.mean(upward_volatilities)) if upward_volatilities else 0.0
+        ),
+        "eval/supra_share_mean": float(np.mean(supra_shares)) if supra_shares else 0.0,
+        "eval/supra_penalty_mean": (
+            float(np.mean(supra_penalties)) if supra_penalties else 0.0
+        ),
+        "eval/agent_prob_mean": float(np.mean(agent_probs)) if agent_probs else 0.0,
+    }
+
+
+def evaluate(
+    agent: Any,
+    env: Any,
+    episodes: int,
+    cfg: Mapping[str, Any] | None = None,
+) -> dict[str, float]:
+    metrics = _evaluate_env(agent, env, episodes)
+    if cfg is None or not bool(cfg.get("robust_eval_enabled", True)):
+        return metrics
+
+    nominal_alpha = float(cfg.get("alpha", 0.0))
+    eval_radius = max(float(cfg.get("robust_radius", 0.0)), 0.15)
+    low_alpha = float(np.clip(nominal_alpha - eval_radius, 0.0, 1.0))
+    high_alpha = float(np.clip(nominal_alpha + eval_radius, 0.0, 1.0))
+    shifted_episodes = max(1, int(np.ceil(int(episodes) / 2)))
+
+    shifted_rows = []
+    for tag, alpha in (
+        ("low", low_alpha),
+        ("nominal", nominal_alpha),
+        ("high", high_alpha),
+    ):
+        eval_cfg = dict(cfg)
+        eval_cfg["alpha"] = float(alpha)
+        shifted_env = make_env(eval_cfg)
+        shifted_metrics = _evaluate_env(agent, shifted_env, shifted_episodes)
+        shifted_env.close()
+        shifted_rows.append((tag, alpha, shifted_metrics))
+
+    metrics["eval/stress_alpha_low"] = low_alpha
+    metrics["eval/stress_alpha_high"] = high_alpha
+    metrics["eval/stress_reward_worst"] = float(
+        min(row[2]["eval/reward_mean"] for row in shifted_rows)
+    )
+    metrics["eval/stress_revenue_worst"] = float(
+        min(row[2]["eval/revenue_mean"] for row in shifted_rows)
+    )
+    metrics["eval/stress_coi_leakage_worst"] = float(
+        max(row[2]["eval/coi_leakage_mean"] for row in shifted_rows)
+    )
+    for tag, alpha, shifted_metrics in shifted_rows:
+        metrics[f"eval/{tag}_alpha"] = float(alpha)
+        metrics[f"eval/{tag}_reward_mean"] = float(shifted_metrics["eval/reward_mean"])
+        metrics[f"eval/{tag}_revenue_mean"] = float(
+            shifted_metrics["eval/revenue_mean"]
+        )
+        metrics[f"eval/{tag}_coi_leakage_mean"] = float(
+            shifted_metrics["eval/coi_leakage_mean"]
+        )
+
+    return metrics

engine/backends/qtable.py

@@ -0,0 +1,139 @@
+from __future__ import annotations
+
+import logging
+import time
+from typing import Any, Mapping
+
+import numpy as np
+
+from .common import evaluate, make_env
+from ..telemetry.wandb import get_wandb_module
+
+logger = logging.getLogger(__name__)
+
+
+def train_qtable(
+    cfg: Mapping[str, Any],
+) -> tuple[object, dict[str, Any]]:
+    from ..lib.discrete import EventQTable
+
+    np.random.seed(int(cfg["seed"]))
+    env = make_env(cfg)
+    eval_env = make_env(cfg)
+    agent = EventQTable(
+        env.action_space.n,
+        int(cfg["n_products"]),
+        (float(cfg["price_low"]), float(cfg["price_high"])),
+        lr=float(cfg["q_lr"]),
+        gamma=float(cfg["gamma"]),
+        n_bins=int(cfg["q_bins"]),
+    )
+
+    total_reward = 0.0
+    total_revenue = 0.0
+    steps = 0
+    epsilon = float(cfg["eps_start"])
+    log_freq = max(1, int(cfg.get("log_freq", 100)))
+    console_progress = bool(cfg.get("console_progress", False))
+    obs, _ = env.reset(seed=int(cfg["seed"]))
+    started_at = time.perf_counter()
+    wandb = get_wandb_module()
+    wandb_live = bool(wandb is not None and wandb.run is not None)
+    step_offset = max(0, int(cfg.get("wandb_step_offset", 0)))
+
+    interval_sums = {
+        "reward": 0.0,
+        "revenue": 0.0,
+        "agent_prob": 0.0,
+        "alpha_adv": 0.0,
+        "coi_leakage": 0.0,
+    }
+    interval_count = 0
+    train_events: list[dict[str, float | int]] = []
+
+    for _ in range(int(cfg["total_timesteps"])):
+        action, state = agent.act(obs, epsilon)
+        nxt, reward, term, trunc, info = env.step(action)
+        done = bool(term or trunc)
+        agent.update(state, action, float(reward), agent.encode(nxt), done)
+
+        total_reward += float(reward)
+        revenue = float(info.get("economics", {}).get("revenue", 0.0))
+        total_revenue += revenue
+        steps += 1
+        interval_sums["reward"] += float(reward)
+        interval_sums["revenue"] += revenue
+        interval_sums["agent_prob"] += float(info.get("agent_prob", 0.0))
+        interval_sums["alpha_adv"] += float(info.get("alpha_adv", 0.0))
+        interval_sums["coi_leakage"] += float(info.get("coi_leakage", 0.0))
+        interval_count += 1
+
+        if steps % log_freq == 0 and interval_count > 0:
+            denom = float(interval_count)
+            event = {
+                "train/reward_mean": interval_sums["reward"] / denom,
+                "train/revenue_mean": interval_sums["revenue"] / denom,
+                "train/agent_prob": interval_sums["agent_prob"] / denom,
+                "train/alpha_adv": interval_sums["alpha_adv"] / denom,
+                "train/coi_leakage": interval_sums["coi_leakage"] / denom,
+                "train/epsilon": float(epsilon),
+                "train/global_step": int(steps),
+            }
+            if wandb_live:
+                try:
+                    wandb.log(dict(event), step=step_offset + int(steps))
+                except Exception:
+                    wandb_live = False
+                    train_events.append(event)
+            else:
+                train_events.append(event)
+            if console_progress:
+                elapsed = max(time.perf_counter() - started_at, 1e-6)
+                speed = steps / elapsed
+                logger.info(
+                    "step=%d/%d reward=%.3f revenue=%.3f eps=%.4f speed=%.1f steps/s",
+                    steps,
+                    int(cfg["total_timesteps"]),
+                    event["train/reward_mean"],
+                    event["train/revenue_mean"],
+                    event["train/epsilon"],
+                    speed,
+                )
+            interval_sums = {key: 0.0 for key in interval_sums}
+            interval_count = 0
+
+        epsilon = max(float(cfg["eps_end"]), epsilon * float(cfg["eps_decay"]))
+        obs = env.reset()[0] if done else nxt
+
+    if interval_count > 0:
+        denom = float(interval_count)
+        tail_event = {
+            "train/reward_mean": interval_sums["reward"] / denom,
+            "train/revenue_mean": interval_sums["revenue"] / denom,
+            "train/agent_prob": interval_sums["agent_prob"] / denom,
+            "train/alpha_adv": interval_sums["alpha_adv"] / denom,
+            "train/coi_leakage": interval_sums["coi_leakage"] / denom,
+            "train/epsilon": float(epsilon),
+            "train/global_step": int(steps),
+        }
+        if wandb_live:
+            try:
+                wandb.log(dict(tail_event), step=step_offset + int(steps))
+            except Exception:
+                wandb_live = False
+                train_events.append(tail_event)
+        else:
+            train_events.append(tail_event)
+
+    metrics: dict[str, Any] = {
+        "train/reward_mean": total_reward / max(steps, 1),
+        "train/revenue_mean": total_revenue / max(steps, 1),
+        "train/epsilon": float(epsilon),
+        "train/global_step": int(cfg["total_timesteps"]),
+    }
+    metrics.update(evaluate(agent, eval_env, int(cfg["eval_episodes"]), cfg=cfg))
+    metrics["_train_events"] = train_events
+
+    env.close()
+    eval_env.close()
+    return agent, metrics

engine/backends/sb3.py

@@ -0,0 +1,217 @@
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from typing import Any, Mapping
+
+from ..lib.callbacks import EvalMetricsCallback, MetricsCallback
+from ..wandb_checkpoint import checkpoint_artifact_name, log_checkpoint_file
+from .common import evaluate, make_env
+
+
+def _net_arch(name: Any) -> list[int]:
+    presets = {
+        "tiny": [32, 32],
+        "small": [64, 64],
+        "medium": [128, 128],
+        "large": [256, 256],
+    }
+    if isinstance(name, (list, tuple)):
+        return [int(v) for v in name]
+    raw = str(name).lower().strip()
+    if raw in presets:
+        return presets[raw]
+    if "x" in raw:
+        try:
+            parsed = [int(v) for v in raw.split("x") if v]
+            return parsed if parsed else presets["small"]
+        except ValueError:
+            return presets["small"]
+    return presets["small"]
+
+
+def _activation(name: Any):
+    try:
+        import torch.nn as nn
+    except ImportError:
+        return None
+    return {
+        "relu": nn.ReLU,
+        "tanh": nn.Tanh,
+        "elu": nn.ELU,
+        "leaky_relu": nn.LeakyReLU,
+    }.get(str(name).lower().strip(), nn.ReLU)
+
+
+def _policy_kwargs(cfg: Mapping[str, Any]) -> dict[str, Any]:
+    kwargs: dict[str, Any] = {"net_arch": _net_arch(cfg.get("arch", "small"))}
+    activation = _activation(cfg.get("activation", "relu"))
+    if activation is not None:
+        kwargs["activation_fn"] = activation
+    return kwargs
+
+
+def build_model(cfg: Mapping[str, Any], env: Any):
+    try:
+        from stable_baselines3 import A2C, DQN, PPO
+    except ImportError as exc:
+        raise ImportError("stable-baselines3 is required for SB3 algorithms") from exc
+
+    algo = str(cfg["algo"])
+    policy_kwargs = _policy_kwargs(cfg)
+    device = str(cfg.get("device", "auto"))
+    seed = int(cfg["seed"])
+
+    if algo == "sac":
+        raise ValueError("sac is not supported with the discrete core env")
+    if algo == "ppo":
+        return PPO(
+            "MlpPolicy",
+            env,
+            verbose=1,
+            device=device,
+            policy_kwargs=policy_kwargs,
+            seed=seed,
+            learning_rate=float(cfg["learning_rate"]),
+            n_steps=int(cfg["n_steps"]),
+            batch_size=int(cfg["batch_size"]),
+            n_epochs=int(cfg["n_epochs"]),
+            gamma=float(cfg["gamma"]),
+            gae_lambda=float(cfg["gae_lambda"]),
+            clip_range=float(cfg["clip_range"]),
+            ent_coef=float(cfg["ent_coef"]),
+        )
+    if algo == "a2c":
+        return A2C(
+            "MlpPolicy",
+            env,
+            verbose=1,
+            device=device,
+            policy_kwargs=policy_kwargs,
+            seed=seed,
+            learning_rate=float(cfg["learning_rate"]),
+            n_steps=max(5, int(cfg["n_steps"]) // 32),
+            gamma=float(cfg["gamma"]),
+            gae_lambda=float(cfg["gae_lambda"]),
+            ent_coef=float(cfg["ent_coef"]),
+        )
+    if algo == "dqn":
+        return DQN(
+            "MlpPolicy",
+            env,
+            verbose=1,
+            device=device,
+            policy_kwargs=policy_kwargs,
+            seed=seed,
+            learning_rate=float(cfg["learning_rate"]),
+            buffer_size=int(cfg["buffer_size"]),
+            batch_size=int(cfg["batch_size"]),
+            gamma=float(cfg["gamma"]),
+            train_freq=int(cfg["train_freq"]),
+            learning_starts=int(cfg["learning_starts"]),
+            target_update_interval=int(cfg["target_update_interval"]),
+            exploration_fraction=float(cfg["exploration_fraction"]),
+            exploration_final_eps=float(cfg["exploration_final_eps"]),
+        )
+    raise ValueError(f"unsupported algo '{algo}'")
+
+
+def train_sb3(cfg: Mapping[str, Any]) -> tuple[object, dict[str, Any]]:
+    try:
+        from stable_baselines3.common.monitor import Monitor
+    except ImportError as exc:
+        raise ImportError("stable-baselines3 is required for SB3 models") from exc
+
+    env = Monitor(make_env(cfg))
+    eval_env = Monitor(make_env(cfg))
+    model = build_model(cfg, env)
+
+    try:
+        import torch
+
+        print(
+            "PHANTOM_DEVICE: "
+            + json.dumps(
+                {
+                    "requested": str(cfg.get("device", "auto")),
+                    "torch_cuda_available": bool(torch.cuda.is_available()),
+                    "torch_device_count": int(torch.cuda.device_count()),
+                    "sb3_device": str(getattr(model, "device", "unknown")),
+                }
+            )
+        )
+    except Exception:
+        pass
+
+    metrics_callback = MetricsCallback(
+        log_histograms=True,
+        log_freq=int(cfg["log_freq"]),
+        hist_freq=int(cfg.get("hist_freq", 500)),
+        step_offset=int(cfg.get("wandb_step_offset", 0)),
+    )
+    eval_callback = EvalMetricsCallback(
+        eval_env,
+        eval_freq=int(cfg["eval_freq"]),
+        n_eval_episodes=int(cfg["eval_episodes"]),
+        step_offset=int(cfg.get("wandb_step_offset", 0)),
+        deterministic=True,
+        verbose=0,
+    )
+    callbacks = [metrics_callback, eval_callback]
+
+    target_steps = int(cfg["total_timesteps"])
+    remaining_steps = max(0, target_steps - int(getattr(model, "num_timesteps", 0)))
+    if remaining_steps > 0:
+        model.learn(
+            total_timesteps=remaining_steps,
+            callback=callbacks,
+            reset_num_timesteps=False,
+        )
+
+    model_dir = Path(str(cfg["model_dir"]))
+    model_dir.mkdir(parents=True, exist_ok=True)
+    model_path = model_dir / f"phantom_{cfg['algo']}"
+    model.save(str(model_path))
+
+    artifact_name = checkpoint_artifact_name(
+        cfg,
+        backend="sb3",
+        sweep_id=os.getenv("WANDB_SWEEP_ID"),
+    )
+    artifact_logged = False
+    try:
+        artifact_logged = bool(
+            log_checkpoint_file(
+                artifact_name,
+                file_path=model_path.with_suffix(".zip"),
+                artifact_file_name="model.zip",
+                metadata={
+                    "algo": str(cfg.get("algo", "ppo")),
+                    "backend": "sb3",
+                    "seed": int(cfg.get("seed", 0)),
+                    "step": int(getattr(model, "num_timesteps", 0)),
+                },
+            )
+        )
+    except Exception:
+        artifact_logged = False
+
+    metrics: dict[str, Any] = evaluate(
+        model,
+        eval_env,
+        int(cfg["eval_episodes"]),
+        cfg=cfg,
+    )
+    metrics["train/global_step"] = int(model.num_timesteps)
+    metrics["model/path"] = str(model_path.with_suffix(".zip"))
+    metrics["model/artifact_name"] = str(artifact_name)
+    metrics["model/artifact_logged"] = float(artifact_logged)
+    metrics["_train_events"] = sorted(
+        [*metrics_callback.events, *eval_callback.events],
+        key=lambda event: int(event.get("train/global_step", 0)),
+    )
+
+    env.close()
+    eval_env.close()
+    return model, metrics

engine/benchmark.py

@@ -0,0 +1,702 @@
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+
+import argparse
+import json
+import logging
+from datetime import datetime, timezone
+from pathlib import Path
+
+# clear stale TPU locks on startup
+if os.path.exists("/dev/accel0"):
+    try:
+        subprocess.run(
+            ["rm", "-f", "/tmp/.libtpu_lockfile", "/tmp/libtpu_lockfile"],
+            stderr=subprocess.DEVNULL,
+        )
+    except:
+        pass
+
+try:
+    import jax
+
+    jax.config.update("jax_threefry_partitionable", True)
+except ImportError:
+    pass
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+
+from .lib.tiers import LinearElasticityPolicy, StaticPolicy, SurgePolicy
+from .logging_utils import configure_logging
+from .spec import TrainSpec
+from .telemetry.wandb import get_wandb_module
+
+wandb = get_wandb_module()
+HAS_WANDB = wandb is not None
+logger = logging.getLogger(__name__)
+
+
+def _log(message: str) -> None:
+    logger.info(message)
+
+
+def _wandb_run_active() -> bool:
+    return bool(HAS_WANDB and getattr(wandb, "run", None) is not None)
+
+
+def _parse_list(raw: str) -> list[str]:
+    return [x.strip().lower() for x in str(raw).split(",") if x.strip()]
+
+
+def _parse_float_list(raw: str) -> list[float]:
+    return [float(x.strip()) for x in str(raw).split(",") if x.strip()]
+
+
+def _truthy(value: str | bool | None) -> bool:
+    if isinstance(value, bool):
+        return value
+    if value is None:
+        return False
+    return str(value).strip().lower() in {"1", "true", "yes", "on"}
+
+
+def _mode_label_from_baseline(is_baseline: bool) -> str:
+    return "baseline" if bool(is_baseline) else "defended"
+
+
+def _action(policy, obs: np.ndarray):
+    out = policy.predict(obs, deterministic=True)
+    action = out[0] if isinstance(out, tuple) else out
+    if isinstance(action, np.ndarray) and action.size == 1:
+        return int(action.reshape(-1)[0])
+    return int(action)
+
+
+def _run_eval_episode(env, policy) -> dict:
+    obs, _ = env.reset()
+    done = False
+    total_reward = 0.0
+    total_revenue = 0.0
+    total_margin = 0.0
+    total_coi = 0.0
+    price_trace: list[float] = []
+    step_count = 0
+
+    while not done:
+        action = _action(policy, obs)
+        obs, reward, term, trunc, info = env.step(action)
+        done = bool(term or trunc)
+        econ = info.get("economics", {})
+        total_reward += float(reward)
+        total_revenue += float(econ.get("revenue", 0.0))
+        total_margin += float(econ.get("margin", 0.0))
+        total_coi += float(econ.get("coi_level", 0.0))
+        prices = np.asarray(info.get("prices", []), dtype=np.float32)
+        if prices.size > 0:
+            price_trace.append(float(np.mean(prices)))
+        step_count += 1
+
+    denom = max(step_count, 1)
+    return {
+        "reward": total_reward,
+        "revenue": total_revenue,
+        "mean_margin": total_margin / denom,
+        "mean_coi": total_coi / denom,
+        "price_trace": price_trace,
+    }
+
+
+def _build_tier(name: str, cfg: dict, alpha: float, *, step_offset: int = 0):
+    from .backends.common import make_env
+
+    tier = name.lower().strip()
+    run_cfg = dict(cfg)
+    run_cfg["alpha"] = float(alpha)
+    run_cfg["wandb_step_offset"] = int(step_offset)
+
+    if tier == "static":
+        return StaticPolicy(int(run_cfg["action_levels"])), []
+
+    if tier == "surge":
+        return (
+            SurgePolicy(
+                n_actions=int(run_cfg["action_levels"]),
+                n_products=int(run_cfg["n_products"]),
+            ),
+            [],
+        )
+
+    if tier == "linear":
+        warmup_env = make_env(run_cfg)
+        policy = LinearElasticityPolicy(
+            n_actions=int(run_cfg["action_levels"]),
+            n_products=int(run_cfg["n_products"]),
+            price_low=float(run_cfg["price_low"]),
+            price_high=float(run_cfg["price_high"]),
+        )
+        policy.fit(
+            warmup_env,
+            warmup_steps=int(run_cfg.get("linear_warmup_steps", 800)),
+            seed=int(run_cfg["seed"]),
+        )
+        warmup_env.close()
+        return policy, []
+
+    if tier == "qtable":
+        from .backends.qtable import train_qtable
+
+        run_cfg["console_progress"] = True
+        agent, metrics = train_qtable(run_cfg)
+        events = metrics.get("_train_events", [])
+        return agent, events if isinstance(events, list) else []
+
+    if tier in {"ppo", "a2c", "dqn"}:
+        from .backends.sb3 import train_sb3
+
+        run_cfg["algo"] = tier
+        agent, metrics = train_sb3(run_cfg)
+        events = metrics.get("_train_events", [])
+        return agent, events if isinstance(events, list) else []
+
+    raise ValueError(f"unsupported tier '{name}'")
+
+
+def _log_train_events(
+    events: list[dict],
+    *,
+    tier_name: str,
+    mode_label: str,
+    alpha: float,
+    step_offset: int,
+) -> int:
+    if not _wandb_run_active():
+        return int(step_offset)
+    if not events:
+        return int(step_offset)
+
+    ordered = sorted(
+        [evt for evt in events if isinstance(evt, dict)],
+        key=lambda evt: int(evt.get("train/global_step", 0)),
+    )
+    if not ordered:
+        return int(step_offset)
+
+    cursor = int(step_offset)
+    for evt in ordered:
+        rel_step = max(1, int(evt.get("train/global_step", 0)))
+        payload = dict(evt)
+        payload.update(
+            {
+                "run.kind": "benchmark",
+                "runtime/backend": tier_name,
+                "study/mode": mode_label,
+                "study/baseline_mode": float(mode_label == "baseline"),
+                "study/alpha": float(alpha),
+            }
+        )
+        try:
+            wandb.log(payload, step=cursor + rel_step)
+        except Exception:
+            return int(step_offset)
+    max_rel = max(max(1, int(evt.get("train/global_step", 0))) for evt in ordered)
+    return cursor + max_rel + 1
+
+
+def run_benchmark(
+    cfg: dict,
+    tiers: list[str],
+    alpha_values: list[float],
+    n_episodes: int,
+    mode_label: str,
+    step_cursor_start: int = 0,
+    eval_alpha_values: list[float] | None = None,
+):
+    from .backends.common import make_env
+
+    rows: list[dict] = []
+    traces: list[dict] = []
+    total_runs = max(1, len(alpha_values) * len(tiers))
+    run_index = 0
+    wandb_step_cursor = int(step_cursor_start)
+
+    for alpha in alpha_values:
+        for tier_name in tiers:
+            run_index += 1
+            _log(
+                f"[{run_index}/{total_runs}] alpha={float(alpha):.2f} tier={tier_name}: training"
+            )
+            policy, train_events = _build_tier(
+                tier_name,
+                cfg,
+                alpha,
+                step_offset=wandb_step_cursor,
+            )
+            prev_cursor = int(wandb_step_cursor)
+            wandb_step_cursor = _log_train_events(
+                train_events,
+                tier_name=tier_name,
+                mode_label=mode_label,
+                alpha=float(alpha),
+                step_offset=wandb_step_cursor,
+            )
+            if wandb_step_cursor == prev_cursor and tier_name in {
+                "qtable",
+                "ppo",
+                "a2c",
+                "dqn",
+            }:
+                wandb_step_cursor += max(1, int(cfg.get("total_timesteps", 1))) + 1
+            eval_targets = (
+                [float(value) for value in eval_alpha_values]
+                if eval_alpha_values
+                else [float(alpha)]
+            )
+            for eval_alpha in eval_targets:
+                env = make_env({**cfg, "alpha": float(eval_alpha)})
+                eps = [_run_eval_episode(env, policy) for _ in range(int(n_episodes))]
+                env.close()
+
+                row = {
+                    "tier": tier_name,
+                    "mode": mode_label,
+                    "alpha": float(eval_alpha),
+                    "train_alpha": float(alpha),
+                    "eval_alpha": float(eval_alpha),
+                    "episodes": int(n_episodes),
+                    "mean_reward": float(np.mean([e["reward"] for e in eps])),
+                    "mean_revenue": float(np.mean([e["revenue"] for e in eps])),
+                    "mean_margin": float(np.mean([e["mean_margin"] for e in eps])),
+                    "mean_coi": float(np.mean([e["mean_coi"] for e in eps])),
+                    "std_revenue": float(np.std([e["revenue"] for e in eps])),
+                }
+                row["objective_score"] = row["mean_reward"]
+                rows.append(row)
+                _log(
+                    f"[{run_index}/{total_runs}] train_alpha={float(alpha):.2f} "
+                    f"eval_alpha={float(eval_alpha):.2f} tier={tier_name}: "
+                    f"reward={row['mean_reward']:.3f} revenue={row['mean_revenue']:.3f} "
+                    f"coi={row['mean_coi']:.4f} score={row['objective_score']:.3f}"
+                )
+
+                max_len = max((len(e["price_trace"]) for e in eps), default=0)
+                step_means = []
+                for step in range(max_len):
+                    vals = [
+                        e["price_trace"][step]
+                        for e in eps
+                        if step < len(e["price_trace"])
+                    ]
+                    step_means.append(float(np.mean(vals)) if vals else np.nan)
+                traces.append(
+                    {
+                        "tier": tier_name,
+                        "alpha": float(eval_alpha),
+                        "train_alpha": float(alpha),
+                        "eval_alpha": float(eval_alpha),
+                        "mean_price_trace": step_means,
+                    }
+                )
+
+                if _wandb_run_active():
+                    try:
+                        wandb.log(
+                            {
+                                "run.kind": "benchmark",
+                                "runtime/backend": tier_name,
+                                "study/mode": mode_label,
+                                "study/baseline_mode": float(mode_label == "baseline"),
+                                "study/alpha": float(eval_alpha),
+                                "study/train_alpha": float(alpha),
+                                "study/eval_alpha": float(eval_alpha),
+                                "eval/reward_mean": row["mean_reward"],
+                                "eval/revenue_mean": row["mean_revenue"],
+                                "eval/margin_mean": row["mean_margin"],
+                                "eval/coi_level_mean": row["mean_coi"],
+                                "objective/score": row["objective_score"],
+                                "objective/coi_preserved": row["mean_coi"],
+                            },
+                            step=wandb_step_cursor,
+                        )
+                    except Exception:
+                        pass
+                    wandb_step_cursor += 1
+
+    return pd.DataFrame(rows), traces, int(wandb_step_cursor)
+
+
+def _plot_outputs(df: pd.DataFrame, traces: list[dict], out_dir: Path, stamp: str):
+    fig1 = plt.figure(figsize=(11, 4.5))
+    if "mode" in df.columns:
+        groups = sorted(df[["tier", "mode"]].drop_duplicates().values.tolist())
+        for tier, mode in groups:
+            sub = df[(df["tier"] == tier) & (df["mode"] == mode)].sort_values("alpha")
+            plt.plot(
+                sub["alpha"],
+                sub["mean_revenue"],
+                marker="o",
+                label=f"{tier}:{mode}",
+            )
+    else:
+        for tier in sorted(df["tier"].unique()):
+            sub = df[df["tier"] == tier].sort_values("alpha")
+            plt.plot(sub["alpha"], sub["mean_revenue"], marker="o", label=tier)
+    plt.xlabel("contamination alpha")
+    plt.ylabel("mean episode revenue")
+    plt.title("Revenue under contamination")
+    plt.grid(alpha=0.3)
+    plt.legend()
+    fig1.tight_layout()
+    rev_path = out_dir / f"benchmark_revenue_{stamp}.png"
+    fig1.savefig(rev_path, dpi=220)
+    plt.close(fig1)
+
+    fig2 = plt.figure(figsize=(11, 4.5))
+    if "mode" in df.columns:
+        groups = sorted(df[["tier", "mode"]].drop_duplicates().values.tolist())
+        for tier, mode in groups:
+            sub = df[(df["tier"] == tier) & (df["mode"] == mode)].sort_values("alpha")
+            plt.plot(
+                sub["alpha"],
+                sub["mean_coi"],
+                marker="s",
+                label=f"{tier}:{mode}",
+            )
+    else:
+        for tier in sorted(df["tier"].unique()):
+            sub = df[df["tier"] == tier].sort_values("alpha")
+            plt.plot(sub["alpha"], sub["mean_coi"], marker="s", label=tier)
+    plt.xlabel("contamination alpha")
+    plt.ylabel("mean COI level")
+    plt.title("COI preservation")
+    plt.grid(alpha=0.3)
+    plt.legend()
+    fig2.tight_layout()
+    coi_path = out_dir / f"benchmark_coi_{stamp}.png"
+    fig2.savefig(coi_path, dpi=220)
+    plt.close(fig2)
+
+    focus_alpha = float(df["alpha"].min()) if not df.empty else 0.0
+    alpha_traces = [t for t in traces if abs(float(t["alpha"]) - focus_alpha) < 1e-9]
+    fig3 = plt.figure(figsize=(11, 4.5))
+    for item in alpha_traces:
+        xs = np.arange(len(item["mean_price_trace"]))
+        ys = np.asarray(item["mean_price_trace"], dtype=np.float32)
+        mode = item.get("mode")
+        label = f"{item['tier']}:{mode}" if mode is not None else str(item["tier"])
+        plt.plot(xs, ys, label=label)
+    plt.xlabel("step")
+    plt.ylabel("mean price")
+    plt.title(f"Price evolution (alpha={focus_alpha:.2f})")
+    plt.grid(alpha=0.3)
+    plt.legend()
+    fig3.tight_layout()
+    price_path = out_dir / f"benchmark_price_trace_{stamp}.png"
+    fig3.savefig(price_path, dpi=220)
+    plt.close(fig3)
+
+    return rev_path, coi_path, price_path
+
+
+def _run_with_args(args, compare_robust_override: bool | None = None):
+    compare_robust = (
+        bool(compare_robust_override)
+        if compare_robust_override is not None
+        else _truthy(os.environ.get("PHANTOM_BENCHMARK_COMPARE_ROBUST"))
+    )
+    baseline_modes = [False, True] if compare_robust else [bool(args.no_robust)]
+
+    base_overrides = {
+        "seed": args.seed,
+        "total_timesteps": args.total_timesteps,
+        "n_products": args.n_products,
+        "N": args.N,
+        "lambda_coi": args.lambda_coi,
+        "robust_radius": args.robust_radius,
+        "robust_points": args.robust_points,
+        "robust_rollouts": args.robust_rollouts,
+        "margin_floor": args.margin_floor,
+        "eta_ux": args.eta_ux,
+        "reward_profit_weight": args.reward_profit_weight,
+        "price_low": args.price_low,
+        "price_high": args.price_high,
+        "action_levels": args.action_levels,
+        "action_scale_low": args.action_scale_low,
+        "action_scale_high": args.action_scale_high,
+        "max_steps": args.max_steps,
+        "learning_rate": args.learning_rate,
+        "batch_size": args.batch_size,
+        "n_steps": args.n_steps,
+        "linear_warmup_steps": args.linear_warmup_steps,
+        "device": args.device,
+    }
+    tiers = _parse_list(args.tiers)
+    alpha_values = _parse_float_list(args.alpha_values)
+    eval_alpha_values = (
+        _parse_float_list(args.eval_alpha_values)
+        if str(getattr(args, "eval_alpha_values", "")).strip()
+        else []
+    )
+    _log(
+        "starting run "
+        + json.dumps(
+            {
+                "tiers": tiers,
+                "alpha_values": alpha_values,
+                "eval_alpha_values": (
+                    eval_alpha_values if eval_alpha_values else alpha_values
+                ),
+                "episodes": int(args.episodes),
+                "total_timesteps": int(args.total_timesteps),
+                "device": str(args.device),
+            }
+        )
+    )
+
+    all_frames: list[pd.DataFrame] = []
+    all_traces: list[dict] = []
+    wandb_step_cursor = 0
+    for baseline_mode in baseline_modes:
+        overrides = dict(base_overrides)
+        overrides["baseline_mode"] = bool(baseline_mode)
+        cfg = TrainSpec.from_flat(
+            {k: v for k, v in overrides.items() if v is not None}
+        ).to_flat_dict()
+        cfg["linear_warmup_steps"] = int(args.linear_warmup_steps)
+        mode_label = _mode_label_from_baseline(bool(baseline_mode))
+        _log(f"mode={mode_label}: begin")
+        df_mode, traces_mode, wandb_step_cursor = run_benchmark(
+            cfg,
+            tiers,
+            alpha_values,
+            args.episodes,
+            mode_label=mode_label,
+            step_cursor_start=wandb_step_cursor,
+            eval_alpha_values=eval_alpha_values,
+        )
+        _log(f"mode={mode_label}: complete ({len(df_mode)} rows)")
+        for trace in traces_mode:
+            trace["mode"] = mode_label
+        all_frames.append(df_mode)
+        all_traces.extend(traces_mode)
+
+    df = pd.concat(all_frames, ignore_index=True) if all_frames else pd.DataFrame()
+    traces = all_traces
+
+    out_dir = Path(args.output_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    stamp = datetime.now(timezone.utc).strftime("%Y%m%d_%H%M%S")
+    csv_path = out_dir / f"benchmark_{stamp}.csv"
+    trace_path = out_dir / f"benchmark_traces_{stamp}.json"
+    df.to_csv(csv_path, index=False)
+    trace_path.write_text(json.dumps(traces, indent=2))
+    rev_path, coi_path, price_path = _plot_outputs(df, traces, out_dir, stamp)
+    _log(f"artifacts written in {out_dir}")
+
+    if not df.empty:
+        best_idx = int(df["objective_score"].idxmax())
+        best = df.iloc[best_idx]
+        _log(
+            "BEST_TIER="
+            + json.dumps(
+                {
+                    "tier": best["tier"],
+                    "mode": best.get("mode", "defended"),
+                    "alpha": float(best["alpha"]),
+                    "objective_score": float(best["objective_score"]),
+                    "mean_revenue": float(best["mean_revenue"]),
+                    "mean_coi": float(best["mean_coi"]),
+                }
+            )
+        )
+    _log(f"BENCHMARK_CSV={csv_path}")
+    _log(f"BENCHMARK_TRACES={trace_path}")
+    _log(f"BENCHMARK_PLOT_REVENUE={rev_path}")
+    _log(f"BENCHMARK_PLOT_COI={coi_path}")
+    _log(f"BENCHMARK_PLOT_PRICE={price_path}")
+
+
+def run_cli(raw_args: list[str] | None = None):
+    configure_logging()
+    parser = argparse.ArgumentParser(description="PHANTOM benchmark orchestrator")
+    parser.add_argument("--project", default="capstone")
+    parser.add_argument("--tiers", default="static,surge,linear,qtable,ppo")
+    parser.add_argument("--alpha-values", default="0.0,0.3,0.6")
+    parser.add_argument("--eval-alpha-values", default="")
+    parser.add_argument("--episodes", type=int, default=10)
+    parser.add_argument("--output-dir", default="engine/studies/results")
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--total-timesteps", type=int, default=25_000)
+    parser.add_argument("--n-products", type=int, default=10)
+    parser.add_argument("--N", type=int, default=100)
+    parser.add_argument("--lambda-coi", type=float, default=0.2)
+    parser.add_argument("--robust-radius", type=float, default=0.15)
+    parser.add_argument("--robust-points", type=int, default=5)
+    parser.add_argument("--robust-rollouts", type=int, default=1)
+    parser.add_argument("--margin-floor", type=float, default=0.85)
+    parser.add_argument("--eta-ux", type=float, default=0.5)
+    parser.add_argument("--reward-profit-weight", type=float, default=1.0)
+    parser.add_argument("--price-low", type=float, default=10.0)
+    parser.add_argument("--price-high", type=float, default=150.0)
+    parser.add_argument("--action-levels", type=int, default=9)
+    parser.add_argument("--action-scale-low", type=float, default=0.8)
+    parser.add_argument("--action-scale-high", type=float, default=1.2)
+    parser.add_argument("--max-steps", type=int, default=100)
+    parser.add_argument("--learning-rate", type=float, default=3e-4)
+    parser.add_argument("--batch-size", type=int, default=256)
+    parser.add_argument("--n-steps", type=int, default=2048)
+    parser.add_argument("--linear-warmup-steps", type=int, default=800)
+    parser.add_argument("--device", type=str, default="auto")
+    parser.add_argument("--no-robust", action="store_true")
+    parser.add_argument("--no-wandb", action="store_true")
+    parser.add_argument("--offline", action="store_true")
+    parser.add_argument("--sweep-agent", action="store_true")
+    parser.add_argument("--sweep-id", type=str)
+    parser.add_argument("--count", type=int, default=0)
+    args = parser.parse_args(raw_args)
+
+    if args.sweep_agent:
+        if args.no_wandb or not HAS_WANDB:
+            raise ValueError("sweep agent requires wandb")
+        if not args.sweep_id:
+            raise ValueError("--sweep-id is required with --sweep-agent")
+
+        def _sweep_run():
+            run = wandb.init(mode="offline" if args.offline else "online")
+            try:
+                key_to_attr = {
+                    "tiers": "tiers",
+                    "alpha_values": "alpha_values",
+                    "eval_alpha_values": "eval_alpha_values",
+                    "episodes": "episodes",
+                    "total_timesteps": "total_timesteps",
+                    "lambda_coi": "lambda_coi",
+                    "robust_radius": "robust_radius",
+                    "robust_points": "robust_points",
+                    "robust_rollouts": "robust_rollouts",
+                    "ambiguity_radius": "robust_radius",
+                    "ambiguity_points": "robust_points",
+                    "ambiguity_rollouts": "robust_rollouts",
+                    "eta_ux": "eta_ux",
+                    "reward_profit_weight": "reward_profit_weight",
+                    "learning_rate": "learning_rate",
+                    "batch_size": "batch_size",
+                    "n_steps": "n_steps",
+                    "baseline_mode": "no_robust",
+                    "no_robust": "no_robust",
+                    "margin_floor": "margin_floor",
+                    "device": "device",
+                }
+                for key in (
+                    "tiers",
+                    "alpha_values",
+                    "eval_alpha_values",
+                    "episodes",
+                    "total_timesteps",
+                    "lambda_coi",
+                    "robust_radius",
+                    "robust_points",
+                    "robust_rollouts",
+                    "ambiguity_radius",
+                    "ambiguity_points",
+                    "ambiguity_rollouts",
+                    "eta_ux",
+                    "reward_profit_weight",
+                    "learning_rate",
+                    "batch_size",
+                    "n_steps",
+                    "baseline_mode",
+                    "no_robust",
+                    "margin_floor",
+                    "device",
+                ):
+                    if key in wandb.config:
+                        setattr(args, key_to_attr[key], wandb.config[key])
+                _run_with_args(args)
+            finally:
+                if run is not None:
+                    wandb.finish()
+
+        wandb.agent(
+            args.sweep_id,
+            function=_sweep_run,
+            count=args.count if args.count > 0 else None,
+        )
+        return
+
+    if args.no_wandb or not HAS_WANDB:
+        _run_with_args(args)
+        return
+
+    tiers = _parse_list(args.tiers)
+    alpha_values = _parse_float_list(args.alpha_values)
+    run_stamp = datetime.now(timezone.utc).strftime("%m%d-%H%M%S")
+    compare_enabled = _truthy(os.environ.get("PHANTOM_BENCHMARK_COMPARE_ROBUST"))
+    compare_tag = "defended-compare" if compare_enabled else "single-mode"
+    modes = (
+        [("baseline", True), ("defended", False)]
+        if compare_enabled
+        else [(_mode_label_from_baseline(bool(args.no_robust)), bool(args.no_robust))]
+    )
+
+    run_idx = 0
+    for tier in tiers:
+        for mode_label, baseline_mode in modes:
+            for alpha in alpha_values:
+                run_idx += 1
+                alpha_token = (
+                    f"{float(alpha):.2f}".rstrip("0").rstrip(".").replace(".", "p")
+                )
+                tier_args = argparse.Namespace(**vars(args))
+                tier_args.tiers = tier
+                tier_args.alpha_values = str(float(alpha))
+                tier_args.no_robust = bool(baseline_mode)
+                run = wandb.init(
+                    project=args.project,
+                    name=(
+                        f"benchmark-{tier}-{mode_label}-a{alpha_token}-{run_stamp}-{run_idx}"
+                    ),
+                    tags=[
+                        "benchmark",
+                        compare_tag,
+                        f"backend:{tier}",
+                        f"mode:{mode_label}",
+                        f"alpha:{alpha_token}",
+                    ],
+                    config={
+                        "run.kind": "benchmark",
+                        "runtime/backend": tier,
+                        "study/mode": mode_label,
+                        "study/baseline_mode": float(baseline_mode),
+                        "study/alpha": float(alpha),
+                        "tiers": tier,
+                        "alpha_values": str(float(alpha)),
+                        "eval_alpha_values": args.eval_alpha_values,
+                        "episodes": args.episodes,
+                        "total_timesteps": args.total_timesteps,
+                        "lambda_coi": args.lambda_coi,
+                        "ambiguity_radius": args.robust_radius,
+                        "ambiguity_points": args.robust_points,
+                        "ambiguity_rollouts": args.robust_rollouts,
+                        "margin_floor": args.margin_floor,
+                        "baseline_mode": float(baseline_mode),
+                        "eta_ux": args.eta_ux,
+                        "reward_profit_weight": args.reward_profit_weight,
+                        "learning_rate": args.learning_rate,
+                        "device": args.device,
+                    },
+                    mode="offline" if args.offline else "online",
+                )
+                try:
+                    _run_with_args(tier_args, compare_robust_override=False)
+                finally:
+                    if run is not None:
+                        wandb.finish()
+
+
+if __name__ == "__main__":
+    run_cli()

engine/engine.py

@@ -0,0 +1,124 @@
+from sys import platform
+import numpy as np
+from .lib.demand import generate_demand_for_actor, estimate_demand
+from .lib.behavior import get_adjusted_transitions, sample_behavior_from_transitions
+from logging import INFO, getLogger
+
+logger = getLogger(__name__)
+logger.setLevel(INFO)
+
+
+class MarketEngine:
+    """implements separate demand distributions for humans and agents per Section 3.1.1"""
+
+    def __init__(
+        self,
+        alpha: float,
+        N: int,
+        human_params: tuple,
+        agent_params: tuple,
+        demand_distribution=np.random.normal,
+        noise_std: float = 1.0,
+        action_weights: dict | None = None,
+    ):
+        # no defaults for D_H, D_A - force explicit experiment design
+        self.alpha = alpha
+        self.N = int(N)
+        self.Nagents = int(N * alpha)
+        self.Nhumans = int(N * (1 - alpha))
+        self.human_params = human_params
+        self.agent_params = agent_params
+        self.noise_std = noise_std
+        self.demand_dist = demand_distribution
+        self.action_weights = action_weights
+
+    def act(self, prices):
+        # generate separate demands d() per actor type
+        demand_h = generate_demand_for_actor(
+            prices,
+            self.human_params,
+            self.noise_std,
+            distribution_method=self.demand_dist,
+        )
+        demand_a = generate_demand_for_actor(
+            prices,
+            self.agent_params,
+            self.noise_std,
+            distribution_method=self.demand_dist,
+        )
+        human_transitions = get_adjusted_transitions(demand_h, human=True)
+        agent_transitions = get_adjusted_transitions(demand_a, human=False)
+        # sample N trajectories in parallel; each chain is independent so threads
+        # do not share state and numpy's per-call RNG is thread-safe
+        human_t = [
+            sample_behavior_from_transitions(human_transitions)
+            for _ in range(self.Nhumans)
+        ]
+        agent_t = [
+            sample_behavior_from_transitions(agent_transitions)
+            for _ in range(self.Nagents)
+        ]
+        # store trajectories for agent probability calculation
+        self.last_trajectories = human_t + agent_t
+
+        demand_proxy = estimate_demand(
+            self.last_trajectories,
+            self.action_weights,
+            normalize=True,
+            per_session=False,
+        )
+        raw_mix = ((1.0 - float(self.alpha)) * demand_h) + (
+            float(self.alpha) * demand_a
+        )
+        total_raw_demand = float(np.sum(raw_mix))
+        if not demand_proxy:
+            return {i: float(raw_mix[i]) for i in range(len(prices))}
+        if total_raw_demand <= 0.0:
+            return {i: 0.0 for i in range(len(prices))}
+        return {
+            i: total_raw_demand * float(demand_proxy.get(i, 0.0)) / 100.0
+            for i in range(len(prices))
+        }
+
+    def measure(self):
+        pass
+
+
+class PricingEngine:
+    def __init__(
+        self,
+    ) -> None:
+        pass
+
+    def act(self, demand):
+        return np.random.uniform(low=25, high=100, size=10)
+
+
+class Limbo:
+    def __init__(self, platform, market) -> None:
+        self.platform_turn = True
+        self.platform = platform
+        self.market = market
+        self.output = None
+
+    def step(self):
+        if self.platform_turn:
+            self.output = self.platform.act(self.output)
+        else:
+            self.output = self.market.act(self.output)
+        self.platform_turn = not self.platform_turn
+        return self.output
+
+    def reset(self):
+        self.platform_turn = True
+        self.output = None
+
+
+if __name__ == "__main__":
+    platform = PricingEngine()
+    market = MarketEngine(
+        alpha=0.3, N=100, human_params=(50, 10), agent_params=(45, 15)
+    )
+    limbo = Limbo(platform, market)
+    for _ in range(10):
+        limbo.step()

engine/jax/__init__.py

@@ -0,0 +1,3 @@
+from .robust import select_adversarial_alpha_jax, _JAX_OK
+
+__all__ = ["select_adversarial_alpha_jax", "_JAX_OK"]

engine/jax/robust.py

@@ -0,0 +1,197 @@
+"""JAX-accelerated robust inner loop for PHANTOM.
+
+provides a drop-in replacement for the sequential alpha-candidate evaluation in
+wrapper.py::_select_adversarial_alpha.  the demand generation and reward
+computation are vmapped over the K candidate alpha values so all candidates are
+evaluated in a single vectorized pass instead of K sequential Python calls.
+
+public surface:
+    select_adversarial_alpha_jax(candidates, prices, human_params, agent_params,
+                                  noise_std, n_sessions, n_products,
+                                  baseline_prices, lambda_coi, info_value,
+                                  reward_profit_weight, rng_key)
+        -> (best_alpha: float, rewards: np.ndarray)
+
+falls back gracefully when JAX is unavailable.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+
+try:
+    import jax
+    import jax.numpy as jnp
+    from jax import vmap, jit
+
+    _JAX_OK = True
+except ImportError:
+    _JAX_OK = False
+
+_JAX_RUNTIME_OK = True
+
+
+def _demand_for_actor_jax(prices, mean, std, noise_std, key):
+    """d(p;theta) = max(0, val - price + noise), normalized to sum 100."""
+    k1, k2 = jax.random.split(key)
+    val = jax.random.normal(k1, shape=prices.shape) * std + mean
+    noise = jax.random.normal(k2, shape=prices.shape) * noise_std
+    demand = jnp.maximum(0.0, val - prices + noise)
+    total = demand.sum()
+    return jnp.where(total > 0, demand / total * 100.0, demand)
+
+
+def _reward_for_candidate(
+    alpha,
+    prices,
+    human_mean,
+    human_std,
+    agent_mean,
+    agent_std,
+    noise_std,
+    baseline_prices,
+    lambda_coi,
+    info_value,
+    reward_profit_weight,
+    key,
+):
+    """compute a scalar reward for a single alpha candidate (pure JAX, vmappable)."""
+    k_h, k_a = jax.random.split(key)
+    # mixed demand proxy: weighted sum of human and agent demand signals
+    demand_h = _demand_for_actor_jax(prices, human_mean, human_std, noise_std, k_h)
+    demand_a = _demand_for_actor_jax(prices, agent_mean, agent_std, noise_std, k_a)
+    demand = (1.0 - alpha) * demand_h + alpha * demand_a
+
+    revenue = jnp.dot(prices, demand)
+    floor_cost = jnp.dot(baseline_prices, demand)
+    profit = revenue - floor_cost
+
+    # agent_prob proxy: use alpha directly (no trajectory available in vectorized path)
+    coi_leakage = alpha * info_value
+    info_budget = jnp.maximum(floor_cost, 1.0)
+    coi_penalty = lambda_coi * coi_leakage * info_budget
+
+    return reward_profit_weight * profit - coi_penalty
+
+
+if _JAX_OK:
+    # compile once; retracing only happens on shape/dtype changes
+    # 12 args: alpha, prices, h_mean, h_std, a_mean, a_std, noise_std,
+    #          baseline_prices, lambda_coi, info_value, reward_profit_weight, key
+    _reward_batched = jit(
+        vmap(
+            _reward_for_candidate,
+            in_axes=(0, None, None, None, None, None, None, None, None, None, None, 0),
+        )
+    )
+
+
+def select_adversarial_alpha_jax(
+    candidates: np.ndarray,
+    prices: np.ndarray,
+    human_params: tuple,
+    agent_params: tuple,
+    noise_std: float,
+    baseline_prices: np.ndarray,
+    lambda_coi: float,
+    info_value: float,
+    reward_profit_weight: float,
+    rng_seed: int = 0,
+) -> tuple[float, np.ndarray]:
+    """evaluate all alpha candidates in a single vmapped pass.
+
+    returns (best_alpha, rewards_array) where best_alpha minimizes reward
+    (worst case for the platform, driving robust policy training).
+
+    falls back to a pure-numpy sequential loop when JAX is unavailable so the
+    wrapper can call this function unconditionally.
+    """
+    global _JAX_RUNTIME_OK
+
+    if not _JAX_OK or not _JAX_RUNTIME_OK:
+        return _fallback(
+            candidates,
+            prices,
+            human_params,
+            agent_params,
+            noise_std,
+            baseline_prices,
+            lambda_coi,
+            info_value,
+            reward_profit_weight,
+        )
+
+    try:
+        k = len(candidates)
+        key = jax.random.PRNGKey(rng_seed)
+        keys = jax.random.split(key, k)
+
+        rewards = np.asarray(
+            _reward_batched(
+                jnp.asarray(candidates, dtype=jnp.float32),
+                jnp.asarray(prices, dtype=jnp.float32),
+                float(human_params[0]),
+                float(human_params[1]),
+                float(agent_params[0]),
+                float(agent_params[1]),
+                float(noise_std),
+                jnp.asarray(baseline_prices, dtype=jnp.float32),
+                float(lambda_coi),
+                float(info_value),
+                float(reward_profit_weight),
+                keys,
+            )
+        )
+        best_idx = int(np.argmin(rewards))
+        return float(candidates[best_idx]), rewards
+    except Exception as exc:
+        # TPU contention / backend init failures can happen in distributed schedulers.
+        # Degrade to numpy path for the remainder of the process.
+        _JAX_RUNTIME_OK = False
+        print(f"PHANTOM_JAX_FALLBACK: {exc}")
+        return _fallback(
+            candidates,
+            prices,
+            human_params,
+            agent_params,
+            noise_std,
+            baseline_prices,
+            lambda_coi,
+            info_value,
+            reward_profit_weight,
+        )
+
+
+def _fallback(
+    candidates,
+    prices,
+    human_params,
+    agent_params,
+    noise_std,
+    baseline_prices,
+    lambda_coi,
+    info_value,
+    reward_profit_weight,
+):
+    """numpy fallback matching the reward formula above."""
+    rewards = []
+    for alpha in candidates:
+        rng = np.random.default_rng()
+        val_h = rng.normal(*human_params, size=len(prices))
+        val_a = rng.normal(*agent_params, size=len(prices))
+        noise_h = rng.normal(0, noise_std, len(prices))
+        noise_a = rng.normal(0, noise_std, len(prices))
+        d_h = np.maximum(0, val_h - prices + noise_h)
+        d_a = np.maximum(0, val_a - prices + noise_a)
+        s_h, s_a = d_h.sum(), d_a.sum()
+        d_h = d_h / s_h * 100 if s_h > 0 else d_h
+        d_a = d_a / s_a * 100 if s_a > 0 else d_a
+        demand = (1.0 - alpha) * d_h + alpha * d_a
+        revenue = float(np.dot(prices, demand))
+        floor_cost = float(np.dot(baseline_prices, demand))
+        profit = revenue - floor_cost
+        coi_penalty = lambda_coi * alpha * info_value * max(floor_cost, 1.0)
+        rewards.append(reward_profit_weight * profit - coi_penalty)
+    rewards = np.array(rewards)
+    best_idx = int(np.argmin(rewards))
+    return float(candidates[best_idx]), rewards

engine/lib/__init__.py

@@ -0,0 +1,38 @@
+from __future__ import annotations
+
+from importlib import import_module
+
+_EXPORTS: dict[str, tuple[str, str]] = {
+    "estimate_demand": (".demand", "estimate_demand"),
+    "estimate_weighted_demand": (".demand", "estimate_weighted_demand"),
+    "generate_demand_for_actor": (".demand", "generate_demand_for_actor"),
+    "sample_behavior": (".behavior", "sample_behavior"),
+    "get_transition_models": (".behavior", "get_transition_models"),
+    "trajectory_to_events": (".behavior", "trajectory_to_events"),
+    "DashboardRenderer": (".render", "DashboardRenderer"),
+    "style_axis": (".render", "style_axis"),
+    "EconomicMetricsWrapper": (".wrappers", "EconomicMetricsWrapper"),
+    "MetricsCallback": (".callbacks", "MetricsCallback"),
+    "EvalMetricsCallback": (".callbacks", "EvalMetricsCallback"),
+    "ProviderBenchmark": (".providers", "ProviderBenchmark"),
+    "ProviderResult": (".providers", "ProviderResult"),
+    "BenchmarkConfig": (".providers", "BenchmarkConfig"),
+    "RandomBaseline": (".providers", "RandomBaseline"),
+    "SurgeBaseline": (".providers", "SurgeBaseline"),
+    "compute_uplift_coi": (".coi", "compute_uplift_coi"),
+    "extract_purchases": (".coi", "extract_purchases"),
+    "compute_agent_probability": (".coi", "compute_agent_probability"),
+    "EventQTable": (".discrete", "EventQTable"),
+}
+
+__all__ = sorted(_EXPORTS)
+
+
+def __getattr__(name: str):
+    if name not in _EXPORTS:
+        raise AttributeError(f"module '{__name__}' has no attribute '{name}'")
+    module_name, attr_name = _EXPORTS[name]
+    module = import_module(module_name, package=__name__)
+    value = getattr(module, attr_name)
+    globals()[name] = value
+    return value

engine/lib/behavior.py

@@ -0,0 +1,190 @@
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parents[2]))
+
+try:
+    from sim.rl.behavior_loader.models import (
+        BehaviorModel,
+        AgentBehaviorModel,
+        aggregate_event_transitions,
+    )
+except ImportError:
+    BehaviorModel = None
+    AgentBehaviorModel = None
+    aggregate_event_transitions = None
+import pandas as pd
+import numpy as np
+from .demand import generate_demand_for_actor
+
+base_dir = Path(__file__).parents[2] / "experiments"
+human_dir = str(base_dir / "collected_data")
+agent_dir = str(base_dir / "agents" / "collected_data")
+
+_cache = {}  # lazy cache for models and base pivots
+# cache keyed by (human: bool, condition_tuple) so we skip Kronecker re-expansion
+# for repeated calls with the same demand condition inside the robustness inner loop
+_transition_cache: dict = {}
+
+
+def _get_base_pivot(human: bool):
+    if (
+        BehaviorModel is None
+        or AgentBehaviorModel is None
+        or aggregate_event_transitions is None
+    ):
+        raise ImportError("behavior loader dependencies are unavailable")
+    key = "human" if human else "agent"
+    if key not in _cache:
+        model = BehaviorModel(human_dir) if human else AgentBehaviorModel(agent_dir)
+        mdp = model.build_MDP()
+        _cache[key] = pd.DataFrame(aggregate_event_transitions(mdp)).fillna(0.0)
+    return _cache[key]
+
+
+def get_transition_models():
+    """load human and agent transition models for agent probability calculation
+
+    returns:
+        tuple: (human_transitions, agent_transitions) as dicts of event->event->prob
+    """
+    if (
+        BehaviorModel is None
+        or AgentBehaviorModel is None
+        or aggregate_event_transitions is None
+    ):
+        raise ImportError("behavior loader dependencies are unavailable")
+
+    human_model = BehaviorModel(human_dir)
+    agent_model = AgentBehaviorModel(agent_dir)
+
+    human_mdp = human_model.build_MDP()
+    agent_mdp = agent_model.build_MDP()
+
+    human_trans = aggregate_event_transitions(human_mdp)
+    agent_trans = aggregate_event_transitions(agent_mdp)
+
+    return human_trans, agent_trans
+
+
+def trajectory_to_events(trajectory: list) -> list:
+    """extract event names from trajectory for KL divergence calculation
+
+    trajectories are in format 'eventName_product0', extract just eventName
+    """
+    return [s.rsplit("_product", 1)[0] if "_product" in s else s for s in trajectory]
+
+
+class _TransitionTable:
+    """numpy-backed transition table; replaces per-step pandas .loc[] indexing.
+
+    the profiling hotspot was DataFrame.xs called ~4-16k times per outer step.
+    converting once to a dense float32 array with an int-keyed state index map
+    reduces each row lookup to a single array slice with no pandas overhead.
+    rows are pre-normalized so sampling requires no per-step division.
+    """
+
+    __slots__ = ("matrix", "states", "state_index", "n_states")
+
+    def __init__(self, df: pd.DataFrame):
+        self.states: list[str] = df.index.tolist()
+        self.state_index: dict[str, int] = {s: i for i, s in enumerate(self.states)}
+        # float64 throughout: float32 row-sums can drift enough to break np.random.choice
+        mat = np.nan_to_num(
+            df.values.astype(np.float64), nan=0.0, posinf=0.0, neginf=0.0
+        )
+        mat = np.clip(mat, 0.0, None)
+        row_sums = mat.sum(axis=1)
+        # dead rows (all zero) get uniform distribution so sampling never receives NaN
+        dead = row_sums <= 0
+        mat[dead] = 1.0
+        row_sums[dead] = float(mat.shape[1])
+        mat = mat / row_sums[:, np.newaxis]
+        # final nan guard in case fp still drifts
+        np.nan_to_num(mat, nan=0.0, copy=False)
+        row_sums2 = mat.sum(axis=1, keepdims=True)
+        row_sums2[row_sums2 <= 0] = 1.0
+        self.matrix: np.ndarray = mat / row_sums2
+        self.n_states: int = len(self.states)
+
+
+def adjust_behavior_to_condition(condition, transition_matrix):
+    # expand NxN transition matrix to (N*P)x(N*P) weighted by demand condition
+    condition = np.asarray(condition, dtype=float)
+    condition = np.nan_to_num(condition, nan=0.0, posinf=0.0, neginf=0.0)
+    condition = np.clip(condition, 0.0, None)
+    s = float(np.sum(condition))
+    cond_norm = (
+        condition / s
+        if np.isfinite(s) and s > 0
+        else np.full(len(condition), 1.0 / max(len(condition), 1), dtype=float)
+    )
+    n_products = len(condition)
+    base_vals = transition_matrix.values
+    base_cols, base_rows = (
+        transition_matrix.columns.tolist(),
+        transition_matrix.index.tolist(),
+    )
+    expanded = np.kron(base_vals, np.outer(cond_norm, cond_norm))
+    new_cols = [f"{c}_product{p}" for c in base_cols for p in range(n_products)]
+    new_rows = [f"{r}_product{p}" for r in base_rows for p in range(n_products)]
+    return pd.DataFrame(expanded, index=new_rows, columns=new_cols)
+
+
+def get_adjusted_transitions(condition, human=True) -> _TransitionTable:
+    """return a _TransitionTable for the given demand condition.
+
+    results are cached by (human, rounded-condition) so that repeated calls with
+    the same condition inside the robustness inner loop (K candidates, same prices)
+    skip the Kronecker expansion entirely.
+    """
+    condition = np.asarray(condition, dtype=float)
+    # round to 4 significant digits for cache key stability
+    cache_key = (human, tuple(np.round(condition, 4).tolist()))
+    if cache_key in _transition_cache:
+        return _transition_cache[cache_key]
+
+    # prevent OOM by capping cache size
+    if len(_transition_cache) > 100:
+        _transition_cache.clear()
+
+    base_pivot = _get_base_pivot(human)
+    df = adjust_behavior_to_condition(condition, base_pivot)
+    table = _TransitionTable(df)
+    _transition_cache[cache_key] = table
+    return table
+
+
+def clear_transition_cache():
+    """drop cached transition tables; call between episodes if condition space is large."""
+    _transition_cache.clear()
+
+
+def sample_behavior_from_transitions(table, max_len=40):
+    """sample a Markov trajectory.
+
+    accepts _TransitionTable (fast path) or a legacy pandas DataFrame so existing
+    call sites that pass a DataFrame directly continue to work unchanged.
+    """
+    if isinstance(table, pd.DataFrame):
+        table = _TransitionTable(table)
+
+    idx = np.random.randint(table.n_states)
+    trajectory = [table.states[idx]]
+    while len(trajectory) < max_len and "checkout" not in trajectory[-1]:
+        row = table.matrix[table.state_index[trajectory[-1]]]
+        idx = int(np.random.choice(table.n_states, p=row))
+        trajectory.append(table.states[idx])
+    return trajectory
+
+
+def sample_behavior(condition, human=True, max_len=40):
+    table = get_adjusted_transitions(condition, human=human)
+    return sample_behavior_from_transitions(table, max_len=max_len)
+
+
+if __name__ == "__main__":
+    t = sample_behavior(generate_demand_for_actor(np.array([10, 20, 30])), human=True)
+    print(t)
+    t = sample_behavior(generate_demand_for_actor(np.array([10, 20, 30])), human=False)
+    print(t)

engine/lib/callbacks.py

@@ -0,0 +1,259 @@
+"""Training callbacks with algorithm-agnostic metric extraction."""
+
+from typing import Any
+
+from stable_baselines3.common.callbacks import BaseCallback, EvalCallback
+import numpy as np
+
+from ..telemetry.wandb import get_wandb_module
+
+
+class MetricsCallback(BaseCallback):
+    """Collects interval train metrics from env info dictionaries."""
+
+    def __init__(
+        self,
+        log_histograms: bool = False,
+        log_freq: int = 100,
+        hist_freq: int = 500,
+        step_offset: int = 0,
+        verbose: int = 0,
+    ):
+        super().__init__(verbose)
+        self.log_histograms = log_histograms
+        self.log_freq = max(1, int(log_freq))
+        self.hist_freq = max(1, int(hist_freq))
+        self.step_offset = max(0, int(step_offset))
+        self._wandb = get_wandb_module()
+        self._wandb_live = bool(self._wandb is not None and self._wandb.run is not None)
+        self._price_samples: list[float] = []
+        self._demand_samples: list[float] = []
+        self._window_sums = {
+            "train/revenue_mean": 0.0,
+            "train/margin_mean": 0.0,
+            "train/coi_level_mean": 0.0,
+            "train/regret_mean": 0.0,
+            "train/profit_mean": 0.0,
+            "train/agent_prob": 0.0,
+            "train/alpha_adv": 0.0,
+            "train/ux_penalty": 0.0,
+            "train/volatility": 0.0,
+            "train/coi_mix": 0.0,
+            "train/coi_base": 0.0,
+            "train/coi_leakage": 0.0,
+            "train/coi_penalty": 0.0,
+        }
+        self._window_count = 0
+        self.events: list[dict[str, Any]] = []
+
+    def _accumulate(self, info: dict[str, Any]) -> None:
+        econ = info.get("economics")
+        if not isinstance(econ, dict):
+            return
+        self._window_sums["train/revenue_mean"] += float(econ.get("revenue", 0.0))
+        self._window_sums["train/margin_mean"] += float(econ.get("margin", 0.0))
+        self._window_sums["train/coi_level_mean"] += float(econ.get("coi_level", 0.0))
+        self._window_sums["train/regret_mean"] += float(econ.get("regret", 0.0))
+        if "profit" in econ:
+            self._window_sums["train/profit_mean"] += float(econ.get("profit", 0.0))
+        if "agent_prob" in econ:
+            self._window_sums["train/agent_prob"] += float(econ.get("agent_prob", 0.0))
+        if "alpha_adv" in econ:
+            self._window_sums["train/alpha_adv"] += float(econ.get("alpha_adv", 0.0))
+        if "ux_penalty" in econ:
+            self._window_sums["train/ux_penalty"] += float(econ.get("ux_penalty", 0.0))
+        if "volatility" in econ:
+            self._window_sums["train/volatility"] += float(econ.get("volatility", 0.0))
+        if "coi_mix" in econ:
+            self._window_sums["train/coi_mix"] += float(econ.get("coi_mix", 0.0))
+        if "coi_base" in econ:
+            self._window_sums["train/coi_base"] += float(econ.get("coi_base", 0.0))
+        if "coi_leakage" in econ:
+            self._window_sums["train/coi_leakage"] += float(
+                econ.get("coi_leakage", 0.0)
+            )
+        if "coi_penalty" in econ:
+            self._window_sums["train/coi_penalty"] += float(
+                econ.get("coi_penalty", 0.0)
+            )
+        self._window_count += 1
+
+    def _accumulate_histograms(self, info: dict[str, Any]) -> None:
+        if not self.log_histograms:
+            return
+
+        for key in ("effective_prices", "prices"):
+            if key not in info:
+                continue
+            try:
+                values = np.asarray(info.get(key), dtype=float).reshape(-1)
+            except Exception:
+                continue
+            if values.size <= 0:
+                continue
+            finite_values = values[np.isfinite(values)]
+            if finite_values.size > 0:
+                self._price_samples.extend(finite_values.tolist())
+            break
+
+        if "demand" in info:
+            try:
+                demand_values = np.asarray(info.get("demand"), dtype=float).reshape(-1)
+            except Exception:
+                demand_values = np.array([], dtype=float)
+            if demand_values.size > 0:
+                finite_demand = demand_values[np.isfinite(demand_values)]
+                if finite_demand.size > 0:
+                    self._demand_samples.extend(finite_demand.tolist())
+
+    def _flush_histograms(self, step: int, force: bool = False) -> None:
+        if not self.log_histograms:
+            return
+        if not force and step % self.hist_freq != 0:
+            return
+        if not self._price_samples and not self._demand_samples:
+            return
+        if self._wandb is None:
+            self._price_samples.clear()
+            self._demand_samples.clear()
+            return
+
+        payload: dict[str, Any] = {}
+        if self._price_samples:
+            payload["train/price_dist"] = self._wandb.Histogram(
+                np.asarray(self._price_samples, dtype=np.float32)
+            )
+        if self._demand_samples:
+            payload["train/demand_dist"] = self._wandb.Histogram(
+                np.asarray(self._demand_samples, dtype=np.float32)
+            )
+
+        if payload and self._wandb_live:
+            try:
+                self._wandb.log(payload, step=self.step_offset + int(step))
+            except Exception:
+                self._wandb_live = False
+
+        self._price_samples.clear()
+        self._demand_samples.clear()
+
+    def _flush(self, step: int, *, force_hist: bool = False) -> None:
+        if self._window_count > 0:
+            denom = float(self._window_count)
+            payload = {
+                key: (value / denom)
+                for key, value in self._window_sums.items()
+                if value != 0.0
+                or key
+                in {
+                    "train/revenue_mean",
+                    "train/margin_mean",
+                    "train/coi_level_mean",
+                    "train/regret_mean",
+                }
+            }
+            payload["train/global_step"] = int(step)
+            if self._wandb_live:
+                try:
+                    self._wandb.log(dict(payload), step=self.step_offset + int(step))
+                except Exception:
+                    self._wandb_live = False
+                    self.events.append(payload)
+            else:
+                self.events.append(payload)
+            for key in self._window_sums:
+                self._window_sums[key] = 0.0
+            self._window_count = 0
+
+        self._flush_histograms(step=step, force=force_hist)
+
+    def _on_step(self) -> bool:
+        for info in self.locals.get("infos", []):
+            if isinstance(info, dict):
+                self._accumulate(info)
+                self._accumulate_histograms(info)
+
+        if self.num_timesteps % self.log_freq == 0:
+            self._flush(step=self.num_timesteps)
+
+        return True
+
+    def _on_training_end(self) -> None:
+        self._flush(step=self.num_timesteps, force_hist=True)
+
+
+class EvalMetricsCallback(EvalCallback):
+    """Deterministic evaluation collector detached from logging backends."""
+
+    def __init__(
+        self,
+        eval_env,
+        eval_freq: int = 1000,
+        n_eval_episodes: int = 5,
+        step_offset: int = 0,
+        **kwargs,
+    ):
+        super().__init__(
+            eval_env, eval_freq=eval_freq, n_eval_episodes=n_eval_episodes, **kwargs
+        )
+        self.step_offset = max(0, int(step_offset))
+        self._wandb = get_wandb_module()
+        self._wandb_live = bool(self._wandb is not None and self._wandb.run is not None)
+        self._eval_stats: dict[str, list[float]] = {
+            "eval/revenue_mean": [],
+            "eval/margin_mean": [],
+            "eval/coi_level_mean": [],
+            "eval/coi_leakage_mean": [],
+            "eval/volatility_mean": [],
+            "eval/agent_prob_mean": [],
+        }
+        self.events: list[dict[str, float | int]] = []
+
+    def _on_step(self) -> bool:
+        result = super()._on_step()
+        if self.n_calls % self.eval_freq == 0 and hasattr(self, "last_mean_reward"):
+            payload: dict[str, float | int] = {
+                "eval/reward_mean": float(self.last_mean_reward),
+                "train/global_step": int(self.num_timesteps),
+            }
+            for key, values in self._eval_stats.items():
+                payload[key] = float(np.mean(values)) if values else 0.0
+
+            if self._wandb_live:
+                try:
+                    self._wandb.log(
+                        dict(payload),
+                        step=self.step_offset + int(self.num_timesteps),
+                    )
+                except Exception:
+                    self._wandb_live = False
+                    self.events.append(payload)
+            else:
+                self.events.append(payload)
+
+            for values in self._eval_stats.values():
+                values.clear()
+
+        return result
+
+    def _log_success_callback(self, locals_: dict, globals_: dict) -> None:
+        # called after each eval episode
+        info = locals_.get("info", {})
+        econ = info.get("economics") if isinstance(info, dict) else None
+        if not isinstance(econ, dict):
+            return
+
+        self._eval_stats["eval/revenue_mean"].append(float(econ.get("revenue", 0.0)))
+        self._eval_stats["eval/margin_mean"].append(float(econ.get("margin", 0.0)))
+        self._eval_stats["eval/coi_level_mean"].append(
+            float(econ.get("coi_level", 0.0))
+        )
+        self._eval_stats["eval/coi_leakage_mean"].append(
+            float(econ.get("coi_leakage", 0.0))
+        )
+        self._eval_stats["eval/volatility_mean"].append(
+            float(econ.get("volatility", 0.0))
+        )
+        self._eval_stats["eval/agent_prob_mean"].append(
+            float(econ.get("agent_prob", 0.0))
+        )

engine/lib/coi.py

@@ -0,0 +1,79 @@
+import numpy as np
+from typing import Dict
+
+
+def compute_agent_probability(
+    trajectory: list,
+    human_transitions: Dict,
+    agent_transitions: Dict,
+    temperature: float = 1.0,
+) -> float:
+    """estimate agent probability via KL divergence between trajectory transitions and reference models
+
+    compares empirical trajectory transition distribution to human/agent prototypes
+
+    args:
+        trajectory: list of state/event strings from session
+        human_transitions: reference transition dict from human MDP (event->event->prob)
+        agent_transitions: reference transition dict from agent MDP (event->event->prob)
+
+    returns:
+        agent probability in [0, 1] via softmax over KL divergences
+    """
+    if len(trajectory) < 2:
+        return 0.0  # insufficient data, assume human
+
+    # build empirical transition distribution from trajectory
+    trans_counts = {}
+    for s, s_next in zip(trajectory[:-1], trajectory[1:]):
+        if s not in trans_counts:
+            trans_counts[s] = {}
+        trans_counts[s][s_next] = trans_counts[s].get(s_next, 0) + 1
+
+    # normalize to probabilities
+    empirical = {}
+    for s, nxt in trans_counts.items():
+        total = sum(nxt.values())
+        empirical[s] = {s_n: cnt / total for s_n, cnt in nxt.items()}
+
+    # compute KL divergence to each prototype
+    def kl_div(p_dist: Dict, q_dist: Dict) -> float:
+        eps = 1e-10
+        # aggregate over all source states in empirical dist
+        kl = 0.0
+        for s in p_dist:
+            if s not in q_dist:
+                continue  # skip states not in reference
+            p_trans, q_trans = p_dist[s], q_dist[s]
+            for k in p_trans:
+                p_val = p_trans[k] + eps
+                q_val = q_trans.get(k, 0.0) + eps
+                kl += p_val * np.log(p_val / q_val)
+        return kl
+
+    kl_human = kl_div(empirical, human_transitions)
+    kl_agent = kl_div(empirical, agent_transitions)
+
+    # convert to probability via softmax (lower KL = higher prob)
+    t = float(max(temperature, 1e-6))
+    exp_h = np.exp(-kl_human / t)
+    exp_a = np.exp(-kl_agent / t)
+    return float(exp_a / (exp_h + exp_a + 1e-10))
+
+
+def extract_purchases(trajectories: list) -> Dict[int, int]:
+    purchases: Dict[int, int] = {}
+    for traj in trajectories:
+        if traj and "checkout" in traj[-1] and "_product" in traj[-1]:
+            prod_id = int(traj[-1].rsplit("_product", 1)[1])
+            purchases[prod_id] = purchases.get(prod_id, 0) + 1
+    return purchases
+
+
+def compute_uplift_coi(
+    prices: np.ndarray, purchases: Dict[int, int], baseline_prices: np.ndarray
+) -> float:
+    # TODO: consider view-weighted fractional purchase for denser signal
+    return float(
+        sum(max(0.0, prices[k] - baseline_prices[k]) * n for k, n in purchases.items())
+    )

engine/lib/demand.py

@@ -0,0 +1,120 @@
+import numpy as np
+
+CATEGORY_WEIGHTS = {"cart": 4.0, "dwell": 2.0, "nav": 1.0, "filter": 0.5}
+ACTION_CATEGORIES = {
+    "cart": {"add_item", "add_to_cart", "remove", "checkout", "purchase"},
+    "dwell": {"hover_title", "hover_paragraph", "hover_link"},
+    "nav": {"page_view", "view_item", "view", "learn_more"},
+    "filter": {"search", "filter_date", "filter_price", "sort"},
+}
+DEFAULT_ACTION_WEIGHTS = {
+    a: CATEGORY_WEIGHTS[c] for c, actions in ACTION_CATEGORIES.items() for a in actions
+}
+
+
+def generate_demand_for_actor(
+    prices: np.ndarray,
+    params: tuple,
+    noise_std: float = 1.0,
+    distribution_method=np.random.normal,
+    normalize: bool = False,
+) -> np.ndarray:
+    """d(p;0) = max(0, valuation - price) + epsi for single actor type
+    params: (mean, std) for valuation distribution D_H or D_A"""
+    val = distribution_method(*params, size=len(prices))
+    noise = distribution_method(0, noise_std, len(prices))
+    demand = np.maximum(0, val - prices + noise)
+    if not normalize:
+        return demand
+    total = np.sum(demand)
+    return demand / total * 100 if total > 0 else demand
+
+
+def estimate_demand(
+    trajectories,
+    action_weights=None,
+    *,
+    normalize: bool = False,
+    per_session: bool = True,
+):
+    return estimate_weighted_demand(
+        trajectories,
+        action_weights,
+        normalize=normalize,
+        per_session=per_session,
+    )
+
+
+def _parse_event_state(state: str):
+    if "_product" not in state:
+        return state, None
+    action, raw_pid = state.rsplit("_product", 1)
+    return action, int(raw_pid) if raw_pid.isdigit() else None
+
+
+def _weight_for_action(action: str, action_weights: dict) -> float:
+    if action in action_weights:
+        return action_weights[action]
+    if action.startswith("hover"):
+        return CATEGORY_WEIGHTS["dwell"]
+    if action.startswith("filter") or action in {"search", "sort"}:
+        return CATEGORY_WEIGHTS["filter"]
+    if action.startswith("add") or action in {"checkout", "purchase", "remove"}:
+        return CATEGORY_WEIGHTS["cart"]
+    return CATEGORY_WEIGHTS["nav"]
+
+
+def estimate_weighted_demand(
+    trajectories,
+    action_weights=None,
+    *,
+    normalize: bool = False,
+    per_session: bool = True,
+):
+    action_weights = (
+        DEFAULT_ACTION_WEIGHTS if action_weights is None else action_weights
+    )
+    scores = {}
+    for traj in trajectories:
+        for state in traj:
+            action, product_id = _parse_event_state(state)
+            if product_id is None:
+                continue
+            w = _weight_for_action(action, action_weights)
+            if w <= 0:
+                continue
+            scores[product_id] = scores.get(product_id, 0.0) + w
+    if not scores:
+        return {}
+
+    if per_session and len(trajectories) > 0:
+        inv_n = 1.0 / float(len(trajectories))
+        scores = {pid: score * inv_n for pid, score in scores.items()}
+
+    if not normalize:
+        return scores
+
+    total = float(sum(scores.values()))
+    if total <= 0:
+        return {}
+    return {pid: (score / total) * 100.0 for pid, score in scores.items()}
+
+
+# Example usage
+if __name__ == "__main__":
+    np.random.seed(42)
+    prices = np.array([20.0, 35.0, 50.0, 65.0])
+    # demo actor-specific demands
+    human_params, agent_params = (50, 10), (45, 15)
+    demand_h = generate_demand_for_actor(prices, human_params)
+    demand_a = generate_demand_for_actor(prices, agent_params)
+    print("Human Demand:", demand_h)
+    print("Agent Demand:", demand_a)
+    from .behavior import sample_behavior
+
+    N, alpha = 200, 0.3
+    n_h, n_a = int(N * (1 - alpha)), int(N * alpha)
+    human_t = [sample_behavior(demand_h, human=True) for _ in range(n_h)]
+    agent_t = [sample_behavior(demand_a, human=False) for _ in range(n_a)]
+    demand_estimate = estimate_demand(human_t + agent_t)
+    print("Estimated Demand from Behavior:", demand_estimate)

engine/lib/discrete.py

@@ -0,0 +1,70 @@
+from collections import defaultdict
+import gymnasium as gym
+from gymnasium import spaces
+import numpy as np
+
+
+class DiscretePriceActionWrapper(gym.ActionWrapper):
+    def __init__(
+        self,
+        env: gym.Env,
+        n_levels: int = 9,
+        min_scale: float = 0.8,
+        max_scale: float = 1.2,
+    ):
+        super().__init__(env)
+        self.scales = np.linspace(min_scale, max_scale, n_levels, dtype=np.float32)
+        self.action_space = spaces.Discrete(n_levels)
+
+    def action(self, action: int):
+        scale = float(self.scales[int(action)])
+        cur = np.asarray(self.env.unwrapped._prices, dtype=np.float32)
+        lo, hi = self.env.unwrapped.price_bounds
+        return np.clip(cur * scale, lo, hi).astype(np.float32)
+
+
+class EventQTable:
+    def __init__(
+        self,
+        n_actions: int,
+        n_products: int,
+        price_bounds: tuple,
+        lr: float = 0.1,
+        gamma: float = 0.99,
+        n_bins: int = 6,
+    ):
+        self.n_actions = int(n_actions)
+        self.n_products = int(n_products)
+        self.lr = float(lr)
+        self.gamma = float(gamma)
+        self.q = defaultdict(lambda: np.zeros(self.n_actions, dtype=np.float32))
+        lo, hi = price_bounds
+        self.demand_bins = np.linspace(0.0, 100.0, n_bins + 1)[1:-1]
+        self.price_bins = np.linspace(lo, hi, n_bins + 1)[1:-1]
+
+    def encode(self, obs: np.ndarray) -> tuple:
+        obs = np.asarray(obs, dtype=np.float32)
+        d = obs[: self.n_products]
+        p = obs[self.n_products : 2 * self.n_products]
+        d_mean = float(np.mean(d)) if d.size else 0.0
+        d_std = float(np.std(d)) if d.size else 0.0
+        p_mean = float(np.mean(p)) if p.size else 0.0
+        return (
+            int(np.digitize(d_mean, self.demand_bins)),
+            int(np.digitize(d_std, self.demand_bins)),
+            int(np.digitize(p_mean, self.price_bins)),
+        )
+
+    def act(self, obs: np.ndarray, eps: float = 0.0) -> tuple[int, tuple]:
+        s = self.encode(obs)
+        if np.random.random() < eps:
+            return int(np.random.randint(self.n_actions)), s
+        return int(np.argmax(self.q[s])), s
+
+    def update(self, s: tuple, a: int, r: float, s2: tuple, done: bool):
+        target = r + (0.0 if done else self.gamma * float(np.max(self.q[s2])))
+        self.q[s][a] += self.lr * (target - self.q[s][a])
+
+    def predict(self, obs: np.ndarray, deterministic: bool = True):
+        a, _ = self.act(obs, 0.0 if deterministic else 0.05)
+        return a, None

engine/lib/providers.py

@@ -0,0 +1,185 @@
+"""Provider benchmarking - compare pricing strategies across contamination levels."""
+
+from dataclasses import dataclass, field
+from typing import Callable, Any
+import numpy as np
+import pandas as pd
+
+try:
+    import wandb
+
+    HAS_WANDB = True
+except ImportError:
+    HAS_WANDB = False
+
+
+class RandomBaseline:
+    """uniform random action selection as a lower-bound baseline"""
+
+    def __init__(self, n_actions: int):
+        self.n = n_actions
+
+    def __call__(self, obs):
+        return int(np.random.randint(self.n))
+
+    def predict(self, obs, **kw):
+        return self(obs), None
+
+
+class SurgeBaseline:
+    """heuristic surge pricing: boost price when demand is above threshold, discount when below.
+    matches the naive pricing rule from thesis Section 3.3.2"""
+
+    def __init__(
+        self, n_actions: int, high_threshold: float = 60.0, low_threshold: float = 30.0
+    ):
+        self.n = n_actions
+        self.mid = n_actions // 2  # identity action (scale ~1.0)
+        self.high_t = high_threshold
+        self.low_t = low_threshold
+
+    def __call__(self, obs):
+        obs = np.asarray(obs, dtype=np.float32)
+        n_prod = len(obs) // 2
+        demand_mean = float(np.mean(obs[:n_prod])) if n_prod > 0 else 0.0
+        if demand_mean >= self.high_t:
+            return min(self.mid + 2, self.n - 1)  # surge: two levels above identity
+        if demand_mean <= self.low_t:
+            return max(self.mid - 2, 0)  # discount: two levels below identity
+        return self.mid  # hold
+
+    def predict(self, obs, **kw):
+        return self(obs), None
+
+
+@dataclass
+class ProviderResult:
+    """Single benchmark result for one provider at one alpha level."""
+
+    name: str
+    alpha: float
+    total_revenue: float
+    mean_revenue: float
+    coi_level: float
+    coi_preserved_pct: float  # vs alpha=0 baseline
+    margin_integrity: float
+    regret: float
+    episodes: int
+
+
+@dataclass
+class BenchmarkConfig:
+    """Configuration for provider benchmark runs."""
+
+    n_episodes: int = 100
+    alpha_range: list[float] = field(default_factory=lambda: [0.0, 0.1, 0.3, 0.5])
+    baseline_name: str = "fixed"
+
+
+class ProviderBenchmark:
+    """Compare pricing providers to prove margin preservation across contamination levels.
+
+    Usage:
+        def env_factory(alpha):
+            return EconomicMetricsWrapper(PHANTOM(alpha=alpha))
+
+        providers = {
+            "fixed": lambda obs: np.ones(10) * 50,
+            "learned": model.predict,
+        }
+
+        benchmark = ProviderBenchmark(env_factory, providers)
+        results = benchmark.run()
+        print(benchmark.summary_table())
+    """
+
+    def __init__(
+        self,
+        env_factory: Callable[[float], Any],
+        providers: dict[str, Callable],
+        config: BenchmarkConfig | None = None,
+    ):
+        self.env_factory = env_factory  # fn(alpha) -> wrapped env
+        self.providers = providers  # {name: fn(obs) -> action}
+        self.config = config or BenchmarkConfig()
+        self.results: list[ProviderResult] = []
+
+    def run(self) -> list[ProviderResult]:
+        """Run benchmark across all providers and alpha levels."""
+        baseline_coi: dict[str, float] = {}  # {provider: coi at alpha=0}
+
+        for alpha in self.config.alpha_range:
+            env = self.env_factory(alpha)
+
+            for name, policy_fn in self.providers.items():
+                revenues, coi_levels, margins = [], [], []
+
+                for _ in range(self.config.n_episodes):
+                    obs, _ = env.reset()
+                    episode_revenue = 0.0
+                    done = False
+
+                    while not done:
+                        action = policy_fn(obs)
+                        # handle sb3 model.predict returning tuple
+                        if isinstance(action, tuple):
+                            action = action[0]
+                        obs, reward, term, trunc, info = env.step(action)
+                        done = term or trunc
+
+                        econ = info.get("economics", {})
+                        episode_revenue += econ.get("revenue", 0)
+                        coi_levels.append(econ.get("coi_level", 0))
+                        margins.append(econ.get("margin", 0))
+
+                    revenues.append(episode_revenue)
+
+                mean_coi = np.mean(coi_levels) if coi_levels else 0.0
+                if alpha == 0.0:
+                    baseline_coi[name] = mean_coi
+
+                base = baseline_coi.get(name, mean_coi)
+                coi_preserved = mean_coi / base if base > 0 else 1.0
+
+                result = ProviderResult(
+                    name=name,
+                    alpha=alpha,
+                    total_revenue=float(np.sum(revenues)),
+                    mean_revenue=float(np.mean(revenues)),
+                    coi_level=mean_coi,
+                    coi_preserved_pct=coi_preserved * 100,
+                    margin_integrity=float(np.mean(margins)) if margins else 0.0,
+                    regret=0.0,  # compute vs optimal if known
+                    episodes=self.config.n_episodes,
+                )
+                self.results.append(result)
+
+                # log to wandb if available
+                if HAS_WANDB and wandb.run is not None:
+                    try:
+                        wandb.log(
+                            {
+                                f"benchmark/{name}/revenue": result.mean_revenue,
+                                f"benchmark/{name}/coi_preserved": result.coi_preserved_pct,
+                                f"benchmark/{name}/margin": result.margin_integrity,
+                                "benchmark/alpha": alpha,
+                            }
+                        )
+                    except Exception:
+                        pass
+
+        return self.results
+
+    def to_dataframe(self) -> pd.DataFrame:
+        """Convert results to pandas DataFrame."""
+        return pd.DataFrame([r.__dict__ for r in self.results])
+
+    def summary_table(self) -> pd.DataFrame:
+        """Pivot table: providers x alpha with revenue/COI metrics."""
+        df = self.to_dataframe()
+        return df.pivot_table(
+            index="name",
+            columns="alpha",
+            values=["mean_revenue", "coi_preserved_pct", "margin_integrity"],
+            aggfunc="mean",
+        )

engine/lib/render.py

@@ -0,0 +1,165 @@
+"""rendering logic for PHANTOM environment dashboard"""
+
+import numpy as np
+import matplotlib.pyplot as plt
+from matplotlib.gridspec import GridSpec
+
+
+def style_axis(ax, title: str = None, xlabel: str = None, ylabel: str = None):
+    ax.spines["top"].set_visible(False)
+    ax.spines["right"].set_visible(False)
+    if title:
+        ax.set_title(title, fontsize=11, fontweight="bold", pad=8)
+    if xlabel:
+        ax.set_xlabel(xlabel, fontsize=9)
+    if ylabel:
+        ax.set_ylabel(ylabel, fontsize=9)
+
+
+class DashboardRenderer:
+    """stateful renderer for PHANTOM market dynamics visualization"""
+
+    def __init__(self):
+        self.fig = None
+        self.gs = None
+
+    def render(self, env) -> None:
+        if self.fig is None:
+            plt.ion()
+            self.fig = plt.figure(figsize=(14, 10))
+            self.gs = GridSpec(
+                3,
+                3,
+                figure=self.fig,
+                hspace=0.35,
+                wspace=0.3,
+                left=0.07,
+                right=0.95,
+                top=0.92,
+                bottom=0.08,
+            )
+            plt.show(block=False)
+
+        self.fig.clear()
+        self.fig.suptitle(
+            f"PHANTOM  Market Dynamics  [t={env._step_count}, a={env.alpha:.2f}]",
+            fontsize=14,
+            fontweight="bold",
+        )
+
+        demand_mat = np.array(env._demand_history).T
+        price_mat = np.array(env._price_history).T
+        elasticity = env._compute_elasticity()
+
+        self._render_scatter(env)
+        self._render_elasticity_bar(env, elasticity)
+        self._render_session_pie(env)
+        self._render_price_heatmap(price_mat)
+        self._render_demand_heatmap(demand_mat)
+        self._render_correlation(env.n_products, price_mat, demand_mat)
+        self._render_revenue(env)
+
+        self.fig.canvas.draw_idle()
+        self.fig.canvas.flush_events()
+
+    def _render_scatter(self, env):
+        ax = self.fig.add_subplot(self.gs[0, 0])
+        prices_flat = np.array(env._price_history).flatten()
+        demands_flat = np.array(env._demand_history).flatten()
+        product_ids = np.tile(np.arange(env.n_products), len(env._price_history))
+        ax.scatter(
+            prices_flat,
+            demands_flat,
+            c=product_ids,
+            cmap="plasma",
+            alpha=0.6,
+            s=15,
+            edgecolors="none",
+        )
+        if len(prices_flat) > 1:
+            z = np.polyfit(prices_flat, demands_flat, 1)
+            p_line = np.linspace(prices_flat.min(), prices_flat.max(), 50)
+            ax.plot(p_line, np.polyval(z, p_line), "--", lw=1.5, alpha=0.8)
+        style_axis(ax, "Price-Demand Relationship", "Price ($)", "Demand")
+
+    def _render_elasticity_bar(self, env, elasticity):
+        ax = self.fig.add_subplot(self.gs[0, 1])
+        ax.barh(range(env.n_products), elasticity, alpha=0.8)
+        ax.axvline(0, lw=0.8, alpha=0.5)
+        ax.axvline(-1, lw=1, ls="--", alpha=0.5)
+        ax.set_yticks(range(env.n_products))
+        ax.set_yticklabels([f"P{i}" for i in range(env.n_products)], fontsize=7)
+        style_axis(ax, "Price Elasticity", "(dQ/dP)(P/Q)", None)
+
+    def _render_session_pie(self, env):
+        ax = self.fig.add_subplot(self.gs[0, 2])
+        n_h, n_a = env.market.Nhumans, env.market.Nagents
+        wedges, _ = ax.pie(
+            [n_h, n_a], startangle=90, wedgeprops={"linewidth": 2, "edgecolor": "white"}
+        )
+        ax.legend(
+            wedges,
+            [f"H ({n_h})", f"A ({n_a})"],
+            loc="lower center",
+            fontsize=8,
+            frameon=False,
+            bbox_to_anchor=(0.5, -0.05),
+        )
+        ax.set_title("Session Mix", fontsize=11, fontweight="bold")
+
+    def _render_price_heatmap(self, price_mat):
+        ax = self.fig.add_subplot(self.gs[1, :2])
+        im = ax.imshow(price_mat, aspect="auto", cmap="viridis", origin="lower")
+        style_axis(ax, "Price Heatmap P(product, t)", "Step", "Product")
+        cbar = self.fig.colorbar(im, ax=ax, fraction=0.03, pad=0.02)
+        cbar.set_label("$", fontsize=8)
+
+    def _render_demand_heatmap(self, demand_mat):
+        ax = self.fig.add_subplot(self.gs[1, 2])
+        im = ax.imshow(demand_mat, aspect="auto", cmap="Blues", origin="lower")
+        style_axis(ax, "Demand Q(product, t)", "Step", None)
+        self.fig.colorbar(im, ax=ax, fraction=0.046, pad=0.02)
+
+    def _render_correlation(self, n_products, price_mat, demand_mat):
+        ax = self.fig.add_subplot(self.gs[2, 0])
+        if price_mat.shape[1] > 2:
+            corr = np.corrcoef(price_mat, demand_mat)[:n_products, n_products:]
+            im = ax.imshow(corr, cmap="RdBu", vmin=-1, vmax=1, aspect="auto")
+            ax.set_xticks(range(n_products))
+            ax.set_yticks(range(n_products))
+            ax.set_xticklabels([f"Q{i}" for i in range(n_products)], fontsize=6)
+            ax.set_yticklabels([f"P{i}" for i in range(n_products)], fontsize=6)
+            self.fig.colorbar(im, ax=ax, fraction=0.046, pad=0.02)
+        style_axis(ax, "Price-Demand Correlation", None, None)
+
+    def _render_revenue(self, env):
+        ax = self.fig.add_subplot(self.gs[2, 1:])
+        n_steps = len(env._revenue_history)
+        demand_std = [np.std(d) for d in env._demand_history]
+        ax.fill_between(range(n_steps), env._revenue_history, alpha=0.3)
+        ax.plot(env._revenue_history, linewidth=2, label="Revenue")
+        ax.set_xlim(0, max(n_steps, 1))
+        ax.set_ylim(0, max(env._revenue_history) * 1.1 if env._revenue_history else 1)
+
+        ax2 = ax.twinx()
+        ax2.plot(
+            range(n_steps),
+            demand_std,
+            linewidth=2,
+            ls="-",
+            alpha=0.9,
+            label="sigma(Demand)",
+        )
+        d_min, d_max = min(demand_std), max(demand_std)
+        margin = (d_max - d_min) * 0.2 if d_max > d_min else 0.5
+        ax2.set_ylim(max(0, d_min - margin), d_max + margin)
+        ax2.set_ylabel("Demand sigma", fontsize=9)
+
+        style_axis(ax, "Revenue & Demand Dispersion", "Step", "Revenue ($)")
+        ax.legend(loc="upper left", fontsize=7, frameon=False)
+        ax2.legend(loc="upper right", fontsize=7, frameon=False)
+
+    def close(self):
+        if self.fig:
+            plt.close(self.fig)
+            self.fig = None

engine/lib/tiers.py

@@ -0,0 +1,101 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Protocol
+
+import numpy as np
+
+
+class PolicyLike(Protocol):
+    def predict(self, obs: np.ndarray, deterministic: bool = True): ...
+
+
+class StaticPolicy:
+    def __init__(self, n_actions: int):
+        self._action = int(max(0, n_actions // 2))
+
+    def predict(self, obs: np.ndarray, deterministic: bool = True):
+        return self._action, None
+
+
+class SurgePolicy:
+    def __init__(
+        self,
+        n_actions: int,
+        n_products: int,
+        high_threshold: float = 60.0,
+        low_threshold: float = 30.0,
+    ):
+        self.n_actions = int(n_actions)
+        self.n_products = int(n_products)
+        self.mid = self.n_actions // 2
+        self.high_t = float(high_threshold)
+        self.low_t = float(low_threshold)
+
+    def predict(self, obs: np.ndarray, deterministic: bool = True):
+        obs_arr = np.asarray(obs, dtype=np.float32)
+        demand = obs_arr[: self.n_products]
+        demand_mean = float(np.mean(demand)) if demand.size > 0 else 0.0
+        if demand_mean >= self.high_t:
+            return min(self.mid + 2, self.n_actions - 1), None
+        if demand_mean <= self.low_t:
+            return max(self.mid - 2, 0), None
+        return self.mid, None
+
+
+@dataclass
+class LinearElasticityPolicy:
+    n_actions: int
+    n_products: int
+    price_low: float
+    price_high: float
+
+    def __post_init__(self):
+        self.n_actions = int(self.n_actions)
+        self.n_products = int(self.n_products)
+        self.price_low = float(self.price_low)
+        self.price_high = float(self.price_high)
+        self._target_price = 0.5 * (self.price_low + self.price_high)
+        self._action_scales = np.linspace(0.8, 1.2, self.n_actions)
+
+    def fit(self, env, warmup_steps: int = 800, seed: int = 42):
+        rng = np.random.default_rng(int(seed))
+        obs, _ = env.reset(seed=int(seed))
+        prices: list[float] = []
+        demands: list[float] = []
+
+        for _ in range(int(max(10, warmup_steps))):
+            action = int(rng.integers(0, self.n_actions))
+            obs, _, term, trunc, info = env.step(action)
+            done = bool(term or trunc)
+
+            p = np.asarray(info.get("prices", []), dtype=np.float32)
+            d = np.asarray(info.get("demand", []), dtype=np.float32)
+            if p.size > 0 and d.size > 0:
+                prices.append(float(np.mean(p)))
+                demands.append(float(np.mean(d)))
+
+            if done:
+                obs, _ = env.reset()
+
+        if len(prices) < 8:
+            self._target_price = 0.5 * (self.price_low + self.price_high)
+            return self
+
+        slope, intercept = np.polyfit(np.asarray(prices), np.asarray(demands), 1)
+        if slope < -1e-6:
+            p_star = -intercept / (2.0 * slope)
+            self._target_price = float(np.clip(p_star, self.price_low, self.price_high))
+        else:
+            self._target_price = 0.5 * (self.price_low + self.price_high)
+        return self
+
+    def predict(self, obs: np.ndarray, deterministic: bool = True):
+        obs_arr = np.asarray(obs, dtype=np.float32)
+        cur_prices = obs_arr[self.n_products : 2 * self.n_products]
+        cur_mean = (
+            float(np.mean(cur_prices)) if cur_prices.size > 0 else self._target_price
+        )
+        scale = self._target_price / max(cur_mean, 1e-6)
+        action = int(np.argmin(np.abs(self._action_scales - scale)))
+        return int(np.clip(action, 0, self.n_actions - 1)), None

engine/lib/wrappers.py

@@ -0,0 +1,104 @@
+"""Economic metrics wrapper - calculates thesis-aligned KPIs and injects into info dict."""
+
+import gymnasium as gym
+import numpy as np
+
+
+class EconomicMetricsWrapper(gym.Wrapper):
+    """Calculates thesis-aligned economic metrics per step, injects into info.
+
+    Metrics follow thesis definitions:
+    - COI level: E[P] - p_min (Definition 1)
+    - Margin: (avg_price - p_min) / avg_price
+    - Regret: 1 - (revenue / baseline_revenue)
+    """
+
+    def __init__(
+        self, env: gym.Env, p_min: float = 10.0, baseline_revenue: float | None = None
+    ):
+        super().__init__(env)
+        self.p_min = p_min
+        self.baseline_revenue = baseline_revenue
+        self._price_history: list[np.ndarray] = []
+        self._revenue_history: list[float] = []
+
+    def reset(self, **kwargs):
+        obs, info = self.env.reset(**kwargs)
+        self._price_history = []
+        self._revenue_history = []
+        return obs, info
+
+    def step(self, action):
+        obs, reward, terminated, truncated, info = self.env.step(action)
+
+        # extract from unwrapped env
+        quoted_prices = np.asarray(self.env.unwrapped._prices, dtype=float)
+        effective_prices = np.asarray(
+            info.get("effective_prices", quoted_prices), dtype=float
+        )
+        if effective_prices.shape != quoted_prices.shape:
+            effective_prices = quoted_prices
+        demand_dict = self.env.unwrapped._demand
+        demand = np.array([demand_dict.get(i, 0.0) for i in range(len(quoted_prices))])
+
+        # core calculations
+        revenue = float(info.get("revenue", np.sum(effective_prices * demand)))
+        quoted_revenue = float(np.sum(quoted_prices * demand))
+        avg_price = float(np.mean(effective_prices))
+        margin = (avg_price - self.p_min) / max(avg_price, 1e-6)
+        coi_level = avg_price - self.p_min  # E[P] - p_min per thesis Def 1
+
+        self._price_history.append(effective_prices.copy())
+        self._revenue_history.append(revenue)
+
+        # regret vs baseline (golden path)
+        regret = 0.0
+        if self.baseline_revenue and self.baseline_revenue > 0:
+            regret = 1.0 - (revenue / self.baseline_revenue)
+
+        # inject structured metrics into info
+        info["economics"] = {
+            "revenue": revenue,
+            "quoted_revenue": quoted_revenue,
+            "margin": margin,
+            "coi_level": coi_level,
+            "regret": regret,
+        }
+        for key in (
+            "coi_mix",
+            "coi_base",
+            "coi_leakage",
+            "coi_penalty",
+            "ux_penalty",
+            "volatility",
+            "upward_volatility",
+            "supra_penalty",
+            "supra_share",
+            "competitive_anchor",
+            "profit",
+            "cost_floor",
+            "reward_revenue",
+            "reward_total",
+            "agent_prob",
+            "alpha_adv",
+            "alpha_nominal",
+            "erosion_share",
+            "effective_price_mean",
+        ):
+            if key in info:
+                info["economics"][key] = info[key]
+        info["prices"] = quoted_prices.copy()
+        info["effective_prices"] = effective_prices.copy()
+        info["demand"] = demand.copy()
+
+        return obs, reward, terminated, truncated, info
+
+    @property
+    def episode_revenue(self) -> float:
+        return sum(self._revenue_history)
+
+    @property
+    def episode_mean_price(self) -> float:
+        if not self._price_history:
+            return 0.0
+        return float(np.mean([np.mean(p) for p in self._price_history]))

engine/logging_utils.py

@@ -0,0 +1,33 @@
+from __future__ import annotations
+
+import logging
+import os
+import sys
+
+_CONFIGURED = False
+
+
+def _resolve_level(raw: str | None) -> int:
+    name = str(raw or os.environ.get("PHANTOM_LOG_LEVEL", "INFO")).upper().strip()
+    return int(getattr(logging, name, logging.INFO))
+
+
+def configure_logging(level: str | None = None) -> None:
+    global _CONFIGURED
+    if _CONFIGURED:
+        return
+
+    logger = logging.getLogger("engine")
+    logger.setLevel(_resolve_level(level))
+    logger.propagate = False
+
+    if logger.handlers:
+        _CONFIGURED = True
+        return
+
+    handler = logging.StreamHandler(stream=sys.stdout)
+    handler.setFormatter(
+        logging.Formatter("%(asctime)s %(levelname)s [%(name)s] %(message)s")
+    )
+    logger.addHandler(handler)
+    _CONFIGURED = True

engine/orchestrators/__init__.py

@@ -0,0 +1,5 @@
+from .benchmark import run_benchmark_cli
+from .sweep_agent import run_sweep_agent
+from .train import run_train_once
+
+__all__ = ["run_benchmark_cli", "run_sweep_agent", "run_train_once"]

engine/orchestrators/benchmark.py

@@ -0,0 +1,7 @@
+from __future__ import annotations
+
+
+def run_benchmark_cli(raw_args: list[str] | None = None) -> None:
+    from ..benchmark import run_cli
+
+    run_cli(raw_args)

engine/orchestrators/sweep_agent.py

@@ -0,0 +1,71 @@
+from __future__ import annotations
+
+from typing import Any, Mapping, Sequence
+
+from ..spec import TrainSpec, run_name
+from ..telemetry.wandb import (
+    current_config,
+    finish_run,
+    get_wandb_module,
+    init_run,
+    run_agent,
+    update_summary,
+)
+from .train import run_with_active_sweep_run
+
+
+def run_sweep_agent(
+    *,
+    project: str,
+    sweep_id: str,
+    count: int,
+    offline: bool,
+    no_wandb: bool,
+    base_overrides: Mapping[str, Any],
+    kind: str,
+    scenario: str,
+    group: str | None,
+    extra_tags: Sequence[str],
+) -> None:
+    if no_wandb:
+        raise ValueError("sweep agent requires wandb")
+    if not sweep_id:
+        raise ValueError("--sweep-id is required with --sweep-agent")
+    if get_wandb_module() is None:
+        raise ImportError("wandb is required for sweep runs")
+
+    mode = "offline" if offline else "online"
+
+    def _sweep_trial() -> None:
+        run = init_run(mode=mode, project=project, group=group, sweep_mode=True)
+        try:
+            merged = dict(base_overrides)
+            merged.update(current_config())
+            spec = TrainSpec.from_flat(merged)
+            if run is not None:
+                run.name = run_name(spec, kind=kind, scenario=scenario)
+            try:
+                run_with_active_sweep_run(
+                    spec,
+                    kind=kind,
+                    scenario=scenario,
+                    group=group,
+                    extra_tags=extra_tags,
+                )
+                update_summary({"run/status": "finished"})
+            except Exception as exc:
+                update_summary(
+                    {
+                        "run/status": "crashed",
+                        "run/error": str(exc),
+                    }
+                )
+                raise
+        finally:
+            finish_run()
+
+    run_agent(
+        sweep_id,
+        _sweep_trial,
+        count=count if count > 0 else None,
+    )

engine/orchestrators/train.py

@@ -0,0 +1,124 @@
+from __future__ import annotations
+
+import json
+from typing import Any, Sequence
+
+from ..spec import TrainSpec, run_metadata, run_name
+from ..telemetry.wandb import (
+    finish_run,
+    get_wandb_module,
+    init_run,
+    log_metrics,
+    update_run_config,
+    update_summary,
+)
+from ..train_core import run_train
+
+
+def _tags_for_run(spec: TrainSpec, kind: str, extra_tags: Sequence[str]) -> list[str]:
+    tags = [
+        kind,
+        spec.algorithm.name,
+        spec.runtime.backend,
+        "baseline" if spec.study.no_robust else "defended",
+    ]
+    tags.extend([tag for tag in extra_tags if tag])
+    return tags
+
+
+def _print_local_metrics(metrics: dict[str, Any]) -> None:
+    print(json.dumps(metrics, indent=2))
+    print("PHANTOM_METRICS:" + json.dumps(metrics))
+
+
+def _log_train_events(events: list[dict[str, Any]], log_freq: int) -> None:
+    if not events:
+        return
+    period = max(1, int(log_freq))
+    last_logged_step = -period
+    for event in sorted(
+        [evt for evt in events if isinstance(evt, dict)],
+        key=lambda evt: int(evt.get("train/global_step", 0)),
+    ):
+        step = int(event.get("train/global_step", 0))
+        if step <= 0 or (step - last_logged_step) < period:
+            continue
+        log_metrics(event, step=step)
+        last_logged_step = step
+
+
+def run_train_once(
+    spec: TrainSpec,
+    *,
+    project: str,
+    offline: bool,
+    no_wandb: bool,
+    kind: str,
+    scenario: str,
+    group: str | None,
+    extra_tags: Sequence[str],
+) -> dict[str, Any]:
+    wandb = get_wandb_module()
+    if no_wandb or wandb is None:
+        result = run_train(spec)
+        _print_local_metrics(result.metrics)
+        return result.metrics
+
+    mode = "offline" if offline else "online"
+    tags = _tags_for_run(spec, kind, extra_tags)
+    metadata = run_metadata(
+        spec,
+        kind=kind,
+        scenario=scenario,
+        group=group,
+        tags=tags,
+    )
+    config = spec.to_flat_dict()
+    config.update(metadata)
+    name = run_name(spec, kind=kind, scenario=scenario)
+    init_run(
+        mode=mode,
+        project=project,
+        config=config,
+        name=name,
+        tags=tags,
+        group=group,
+        sweep_mode=False,
+    )
+
+    try:
+        result = run_train(spec)
+        _log_train_events(result.events, spec.runtime.log_freq)
+        metrics = result.metrics
+        step = int(metrics.get("train/global_step", spec.runtime.total_timesteps))
+        log_metrics(metrics, step=step)
+        update_summary(metrics)
+        return metrics
+    finally:
+        finish_run()
+
+
+def run_with_active_sweep_run(
+    spec: TrainSpec,
+    *,
+    kind: str,
+    scenario: str,
+    group: str | None,
+    extra_tags: Sequence[str],
+) -> dict[str, Any]:
+    tags = _tags_for_run(spec, kind, extra_tags)
+    metadata = run_metadata(
+        spec,
+        kind=kind,
+        scenario=scenario,
+        group=group,
+        tags=tags,
+    )
+    update_run_config({**spec.to_flat_dict(), **metadata})
+    result = run_train(spec)
+    _log_train_events(result.events, spec.runtime.log_freq)
+    metrics = result.metrics
+    step = int(metrics.get("train/global_step", spec.runtime.total_timesteps))
+    log_metrics(metrics, step=step)
+    update_summary(metrics)
+    return metrics

engine/project.json

@@ -0,0 +1,138 @@
+{
+  "$schema": "../node_modules/nx/schemas/project-schema.json",
+  "name": "research",
+  "projectType": "application",
+  "sourceRoot": "engine",
+  "targets": {
+    "install": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash scripts/nx_research.sh install",
+        "cwd": "."
+      }
+    },
+    "test": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": ".venv/bin/pytest -v",
+        "cwd": "."
+      }
+    },
+    "train": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "bash scripts/nx_research.sh train",
+        "cwd": "."
+      }
+    },
+    "benchmark": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "bash scripts/nx_research.sh benchmark",
+        "cwd": "."
+      }
+    },
+    "benchmark-simple": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "bash scripts/nx_research.sh benchmark-simple",
+        "cwd": "."
+      }
+    },
+    "benchmark-agent": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "bash scripts/nx_research.sh benchmark-agent",
+        "cwd": "."
+      }
+    },
+    "train-agent": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "bash scripts/nx_research.sh train-agent",
+        "cwd": "."
+      }
+    },
+    "train-bootstrap": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash scripts/nx_research.sh train-bootstrap",
+        "cwd": "."
+      }
+    },
+    "stats": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash scripts/nx_research.sh stats",
+        "cwd": "."
+      }
+    },
+    "docker-train-publish": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash scripts/nx_research.sh docker-train-publish",
+        "cwd": "."
+      }
+    },
+    "whoclicked-publish": {
+      "executor": "nx:run-commands",
+      "dependsOn": [
+        "install"
+      ],
+      "options": {
+        "command": "bash scripts/nx_research.sh whoclicked-publish",
+        "cwd": "."
+      }
+    },
+    "tpu-ray-bootstrap": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash scripts/nx_research.sh tpu-ray-bootstrap",
+        "cwd": "."
+      }
+    },
+    "tpu-ray-deps": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash scripts/nx_research.sh tpu-ray-deps",
+        "cwd": "."
+      }
+    },
+    "tpu-ray-verify": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash scripts/nx_research.sh tpu-ray-verify",
+        "cwd": "."
+      }
+    },
+    "tpu-ray-teardown": {
+      "executor": "nx:run-commands",
+      "options": {
+        "command": "bash scripts/nx_research.sh tpu-ray-teardown",
+        "cwd": "."
+      }
+    }
+  },
+  "tags": [
+    "scope:research",
+    "type:python"
+  ]
+}

engine/spec.py

@@ -0,0 +1,353 @@
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+import os
+from typing import Any, Mapping, Sequence
+
+
+def _truthy(value: str | bool | None) -> bool:
+    if isinstance(value, bool):
+        return value
+    if value is None:
+        return False
+    return str(value).strip().lower() in {"1", "true", "yes", "on"}
+
+
+def _normalize_keys(raw: Mapping[str, Any]) -> dict[str, Any]:
+    alias_map = {
+        "algorithm": "algo",
+        "algorithm.name": "algo",
+        "env.n_products": "n_products",
+        "env.action_levels": "action_levels",
+        "env.action_scale_low": "action_scale_low",
+        "env.action_scale_high": "action_scale_high",
+        "env.price_low": "price_low",
+        "env.price_high": "price_high",
+        "env.max_steps": "max_steps",
+        "env.margin_floor": "margin_floor",
+        "env.margin_floor_patience": "margin_floor_patience",
+        "env.n_sessions": "N",
+        "study.alpha": "alpha",
+        "study.lambda_coi": "lambda_coi",
+        "study.robust_radius": "robust_radius",
+        "study.robust_points": "robust_points",
+        "study.robust_rollouts": "robust_rollouts",
+        "study.ambiguity_radius": "robust_radius",
+        "study.ambiguity_points": "robust_points",
+        "study.ambiguity_rollouts": "robust_rollouts",
+        "study.info_value": "info_value",
+        "study.eta_ux": "eta_ux",
+        "study.reward_profit_weight": "reward_profit_weight",
+        "ambiguity_radius": "robust_radius",
+        "ambiguity_points": "robust_points",
+        "ambiguity_rollouts": "robust_rollouts",
+        "baseline_mode": "no_robust",
+        "stress_eval_enabled": "robust_eval_enabled",
+        "optimizer.learning_rate": "learning_rate",
+        "optimizer.gamma": "gamma",
+        "optimizer.batch_size": "batch_size",
+        "optimizer.n_steps": "n_steps",
+        "runtime.backend": "backend",
+        "runtime.device": "device",
+        "runtime.seed": "seed",
+        "runtime.total_timesteps": "total_timesteps",
+        "runtime.checkpoint_interval": "checkpoint_interval",
+        "runtime.hist_freq": "hist_freq",
+        "eval.eval_freq": "eval_freq",
+        "eval.eval_episodes": "eval_episodes",
+    }
+    normalized: dict[str, Any] = {}
+    for key, value in raw.items():
+        canonical = alias_map.get(str(key), str(key))
+        normalized[canonical] = value
+    return normalized
+
+
+@dataclass(frozen=True)
+class AlgorithmSpec:
+    name: str = "ppo"
+
+
+@dataclass(frozen=True)
+class EnvSpec:
+    n_products: int = 10
+    n_sessions: int = 100
+    price_low: float = 10.0
+    price_high: float = 150.0
+    action_levels: int = 9
+    action_scale_low: float = 0.8
+    action_scale_high: float = 1.2
+    max_steps: int = 100
+    margin_floor: float = 0.05
+    margin_floor_patience: int = 5
+    agent_mu: float = 45.0
+    agent_std: float = 15.0
+
+
+@dataclass(frozen=True)
+class StudySpec:
+    alpha: float = 0.3
+    lambda_coi: float = 0.2
+    robust_radius: float = 0.15
+    robust_points: int = 5
+    robust_rollouts: int = 1
+    info_value: float = 1.0
+    eta_ux: float = 0.5
+    reward_profit_weight: float = 1.0
+    no_robust: bool = False
+
+
+@dataclass(frozen=True)
+class OptimizerSpec:
+    learning_rate: float = 3e-4
+    gamma: float = 0.99
+    buffer_size: int = 50_000
+    batch_size: int = 256
+    tau: float = 0.005
+    train_freq: int = 1
+    learning_starts: int = 1_000
+    target_update_interval: int = 1_000
+    exploration_fraction: float = 0.2
+    exploration_final_eps: float = 0.05
+    n_steps: int = 2_048
+    n_epochs: int = 10
+    gae_lambda: float = 0.95
+    clip_range: float = 0.2
+    ent_coef: float = 0.0
+    q_lr: float = 0.1
+    q_bins: int = 6
+    eps_start: float = 1.0
+    eps_end: float = 0.05
+    eps_decay: float = 0.9995
+    arch: str = "small"
+    activation: str = "relu"
+    vf_coef: float = 0.5
+    max_grad_norm: float = 0.5
+
+
+@dataclass(frozen=True)
+class RuntimeSpec:
+    project: str = "capstone"
+    backend: str = "sb3"
+    device: str = "auto"
+    seed: int = 42
+    total_timesteps: int = 50_000
+    checkpoint_interval: int = 200_000
+    model_dir: str = "engine/models"
+    log_freq: int = 100
+    hist_freq: int = 500
+
+
+@dataclass(frozen=True)
+class EvalSpec:
+    eval_freq: int = 1_000
+    eval_episodes: int = 5
+    robust_eval_enabled: bool = True
+
+
+@dataclass(frozen=True)
+class TrainSpec:
+    algorithm: AlgorithmSpec = field(default_factory=AlgorithmSpec)
+    env: EnvSpec = field(default_factory=EnvSpec)
+    study: StudySpec = field(default_factory=StudySpec)
+    optimizer: OptimizerSpec = field(default_factory=OptimizerSpec)
+    runtime: RuntimeSpec = field(default_factory=RuntimeSpec)
+    eval: EvalSpec = field(default_factory=EvalSpec)
+
+    def to_flat_dict(self) -> dict[str, Any]:
+        return {
+            "project": self.runtime.project,
+            "algo": self.algorithm.name,
+            "seed": self.runtime.seed,
+            "total_timesteps": self.runtime.total_timesteps,
+            "eval_episodes": self.eval.eval_episodes,
+            "eval_freq": self.eval.eval_freq,
+            "log_freq": self.runtime.log_freq,
+            "model_dir": self.runtime.model_dir,
+            "backend": self.runtime.backend,
+            "device": self.runtime.device,
+            "checkpoint_interval": self.runtime.checkpoint_interval,
+            "hist_freq": self.runtime.hist_freq,
+            "n_products": self.env.n_products,
+            "N": self.env.n_sessions,
+            "price_low": self.env.price_low,
+            "price_high": self.env.price_high,
+            "action_levels": self.env.action_levels,
+            "action_scale_low": self.env.action_scale_low,
+            "action_scale_high": self.env.action_scale_high,
+            "max_steps": self.env.max_steps,
+            "margin_floor": self.env.margin_floor,
+            "margin_floor_patience": self.env.margin_floor_patience,
+            "agent_mu": self.env.agent_mu,
+            "agent_std": self.env.agent_std,
+            "alpha": self.study.alpha,
+            "lambda_coi": self.study.lambda_coi,
+            "robust_radius": self.study.robust_radius,
+            "robust_points": self.study.robust_points,
+            "robust_rollouts": self.study.robust_rollouts,
+            "info_value": self.study.info_value,
+            "eta_ux": self.study.eta_ux,
+            "reward_profit_weight": self.study.reward_profit_weight,
+            "no_robust": self.study.no_robust,
+            "learning_rate": self.optimizer.learning_rate,
+            "gamma": self.optimizer.gamma,
+            "buffer_size": self.optimizer.buffer_size,
+            "batch_size": self.optimizer.batch_size,
+            "tau": self.optimizer.tau,
+            "train_freq": self.optimizer.train_freq,
+            "learning_starts": self.optimizer.learning_starts,
+            "target_update_interval": self.optimizer.target_update_interval,
+            "exploration_fraction": self.optimizer.exploration_fraction,
+            "exploration_final_eps": self.optimizer.exploration_final_eps,
+            "n_steps": self.optimizer.n_steps,
+            "n_epochs": self.optimizer.n_epochs,
+            "gae_lambda": self.optimizer.gae_lambda,
+            "clip_range": self.optimizer.clip_range,
+            "ent_coef": self.optimizer.ent_coef,
+            "q_lr": self.optimizer.q_lr,
+            "q_bins": self.optimizer.q_bins,
+            "eps_start": self.optimizer.eps_start,
+            "eps_end": self.optimizer.eps_end,
+            "eps_decay": self.optimizer.eps_decay,
+            "arch": self.optimizer.arch,
+            "activation": self.optimizer.activation,
+            "vf_coef": self.optimizer.vf_coef,
+            "max_grad_norm": self.optimizer.max_grad_norm,
+            "robust_eval_enabled": self.eval.robust_eval_enabled,
+        }
+
+    @classmethod
+    def from_flat(
+        cls,
+        raw: Mapping[str, Any] | None = None,
+        *,
+        env_vars: Mapping[str, str] | None = None,
+    ) -> "TrainSpec":
+        base = cls().to_flat_dict()
+        incoming = _normalize_keys(raw or {})
+        base.update({k: v for k, v in incoming.items() if v is not None})
+
+        runtime_env = os.environ if env_vars is None else env_vars
+        base["device"] = str(
+            base.get("device", runtime_env.get("PHANTOM_DEVICE", "auto"))
+        )
+
+        backend = str(base.get("backend", "sb3")).lower()
+        if backend == "auto":
+            backend = "sb3"
+        if backend != "sb3":
+            backend = "sb3"
+
+        no_robust = _truthy(base.get("no_robust"))
+        if no_robust:
+            base["lambda_coi"] = 0.0
+            base["robust_radius"] = 0.0
+            base["robust_points"] = 1
+            base["robust_rollouts"] = 1
+
+        return cls(
+            algorithm=AlgorithmSpec(name=str(base["algo"]).lower().strip()),
+            env=EnvSpec(
+                n_products=int(base["n_products"]),
+                n_sessions=int(base["N"]),
+                price_low=float(base["price_low"]),
+                price_high=float(base["price_high"]),
+                action_levels=int(base["action_levels"]),
+                action_scale_low=float(base["action_scale_low"]),
+                action_scale_high=float(base["action_scale_high"]),
+                max_steps=int(base["max_steps"]),
+                margin_floor=float(base["margin_floor"]),
+                margin_floor_patience=int(base["margin_floor_patience"]),
+                agent_mu=float(base.get("agent_mu", 45.0)),
+                agent_std=float(base.get("agent_std", 15.0)),
+            ),
+            study=StudySpec(
+                alpha=float(base["alpha"]),
+                lambda_coi=float(base["lambda_coi"]),
+                robust_radius=float(base["robust_radius"]),
+                robust_points=int(base["robust_points"]),
+                robust_rollouts=int(base["robust_rollouts"]),
+                info_value=float(base["info_value"]),
+                eta_ux=float(base["eta_ux"]),
+                reward_profit_weight=float(base["reward_profit_weight"]),
+                no_robust=no_robust,
+            ),
+            optimizer=OptimizerSpec(
+                learning_rate=float(base["learning_rate"]),
+                gamma=float(base["gamma"]),
+                buffer_size=int(base["buffer_size"]),
+                batch_size=int(base["batch_size"]),
+                tau=float(base["tau"]),
+                train_freq=int(base["train_freq"]),
+                learning_starts=int(base["learning_starts"]),
+                target_update_interval=int(base["target_update_interval"]),
+                exploration_fraction=float(base["exploration_fraction"]),
+                exploration_final_eps=float(base["exploration_final_eps"]),
+                n_steps=int(base["n_steps"]),
+                n_epochs=int(base["n_epochs"]),
+                gae_lambda=float(base["gae_lambda"]),
+                clip_range=float(base["clip_range"]),
+                ent_coef=float(base["ent_coef"]),
+                q_lr=float(base["q_lr"]),
+                q_bins=int(base["q_bins"]),
+                eps_start=float(base["eps_start"]),
+                eps_end=float(base["eps_end"]),
+                eps_decay=float(base["eps_decay"]),
+                arch=str(base["arch"]),
+                activation=str(base["activation"]),
+                vf_coef=float(base["vf_coef"]),
+                max_grad_norm=float(base["max_grad_norm"]),
+            ),
+            runtime=RuntimeSpec(
+                project=str(base["project"]),
+                backend=backend,
+                device=str(base["device"]),
+                seed=int(base["seed"]),
+                total_timesteps=int(base["total_timesteps"]),
+                checkpoint_interval=int(base["checkpoint_interval"]),
+                model_dir=str(base["model_dir"]),
+                log_freq=int(base["log_freq"]),
+                hist_freq=int(base["hist_freq"]),
+            ),
+            eval=EvalSpec(
+                eval_freq=int(base["eval_freq"]),
+                eval_episodes=int(base["eval_episodes"]),
+                robust_eval_enabled=_truthy(base.get("robust_eval_enabled", True)),
+            ),
+        )
+
+
+def run_name(spec: TrainSpec, *, kind: str, scenario: str) -> str:
+    alpha_token = f"{float(spec.study.alpha):.2f}".rstrip("0").rstrip(".")
+    mode = "baseline" if bool(spec.study.no_robust) else "defended"
+    return (
+        f"{kind}/{spec.algorithm.name}/{spec.runtime.backend}/"
+        f"{spec.runtime.device}/{scenario}/a{alpha_token}/{mode}/s{spec.runtime.seed}"
+    )
+
+
+def run_metadata(
+    spec: TrainSpec,
+    *,
+    kind: str,
+    scenario: str,
+    group: str | None = None,
+    tags: Sequence[str] = (),
+) -> dict[str, Any]:
+    mode = "baseline" if bool(spec.study.no_robust) else "defended"
+    metadata: dict[str, Any] = {
+        "run.kind": str(kind),
+        "run.algo": spec.algorithm.name,
+        "run.backend": spec.runtime.backend,
+        "run.device": spec.runtime.device,
+        "run.scenario": str(scenario),
+        "run.seed": spec.runtime.seed,
+        "run.tags": list(tags),
+        "study/alpha": float(spec.study.alpha),
+        "study/mode": mode,
+        "study/baseline_mode": float(bool(spec.study.no_robust)),
+        "tiers": spec.algorithm.name,
+    }
+    if group:
+        metadata["run.group"] = group
+    return metadata

engine/studies/factors.py

@@ -0,0 +1,33 @@
+"""shared factor definitions for experimental designs"""
+import numpy as np
+from dataclasses import dataclass
+
+@dataclass
+class Factor:
+    name: str
+    levels: list
+    primary: bool = True  # full cross vs sampled
+
+# demand functions with compatible signatures
+def demand_linear(mu, sigma, size): return np.maximum(0, np.random.normal(mu, sigma, size))
+def demand_uniform(mu, sigma, size): return np.random.uniform(mu - sigma, mu + sigma, size)
+def demand_exponential(mu, sigma, size): return np.random.exponential(mu, size)
+def demand_logistic(mu, sigma, size): return np.random.logistic(mu, sigma, size)
+
+DEMAND_FUNCTIONS = {
+    "linear": demand_linear,
+    "uniform": demand_uniform,
+    "exponential": demand_exponential,
+    "logistic": demand_logistic,
+}
+
+FACTORS = [
+    Factor("demand_fn", list(DEMAND_FUNCTIONS.keys()), primary=True),
+    Factor("alpha", [0.1, 0.3, 0.5, 0.7], primary=True),
+    Factor("n_products", [5, 15, 30, 50], primary=True),
+    Factor("demand_mu", [30.0, 50.0, 70.0], primary=False),
+    Factor("demand_sigma", [5.0, 10.0, 20.0], primary=False),
+    Factor("N", [100, 500, 1000], primary=False),
+]
+
+SEEDS_PER_CONFIG = 5

engine/studies/full_factorial.py

@@ -0,0 +1,104 @@
+"""full factorial design - all factor combinations"""
+
+import sys
+
+sys.path.insert(0, "..")
+import logging
+from itertools import product
+import json
+import hashlib
+from pathlib import Path
+from concurrent.futures import ProcessPoolExecutor
+from .factors import FACTORS, DEMAND_FUNCTIONS, SEEDS_PER_CONFIG
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+log = logging.getLogger(__name__)
+
+
+def generate_configs():
+    """generate all factor combinations with seeds"""
+    all_levels = [f.levels for f in FACTORS]
+    names = [f.name for f in FACTORS]
+
+    configs = []
+    for combo in product(*all_levels):
+        base = {names[i]: combo[i] for i in range(len(names))}
+        for seed in range(SEEDS_PER_CONFIG):
+            cfg = {**base, "seed": seed}
+            cfg["id"] = hashlib.md5(
+                json.dumps(cfg, sort_keys=True).encode()
+            ).hexdigest()[:8]
+            configs.append(cfg)
+    return configs
+
+
+def run_single(cfg: dict) -> dict:
+    """execute one experiment config, return metrics"""
+    from engine.wrapper import PHANTOM
+    import numpy as np
+
+    np.random.seed(cfg["seed"])
+    demand_fn = DEMAND_FUNCTIONS[cfg["demand_fn"]]
+
+    env = PHANTOM(
+        n_products=cfg["n_products"],
+        alpha=cfg["alpha"],
+        N=cfg["N"],
+    )
+    env.market.demand = (demand_fn, (cfg["demand_mu"], cfg["demand_sigma"]))
+
+    obs, _ = env.reset()
+    total_reward, steps = 0.0, 0
+
+    for _ in range(100):
+        action = env.action_space.sample()
+        obs, reward, term, trunc, _ = env.step(action)
+        total_reward += reward
+        steps += 1
+        if term:
+            break
+
+    env.close()
+    return {
+        "id": cfg["id"],
+        "config": cfg,
+        "total_reward": total_reward,
+        "avg_reward": total_reward / steps if steps > 0 else 0.0,
+        "steps": steps,
+    }
+
+
+def run_study(max_workers: int = None, output: str = "results_full.jsonl"):
+    configs = generate_configs()
+    log.info(
+        f"full factorial: {len(configs)} configs ({len(configs) // SEEDS_PER_CONFIG} unique × {SEEDS_PER_CONFIG} seeds)"
+    )
+
+    results = []
+    with ProcessPoolExecutor(max_workers=max_workers) as ex:
+        for i, result in enumerate(ex.map(run_single, configs)):
+            results.append(result)
+            if (i + 1) % 100 == 0:
+                log.info(f"progress: {i + 1}/{len(configs)}")
+
+    Path(output).write_text("\n".join(json.dumps(r) for r in results))
+    log.info(f"wrote {len(results)} results to {output}")
+    return results
+
+
+if __name__ == "__main__":
+    import argparse
+
+    p = argparse.ArgumentParser()
+    p.add_argument("--workers", type=int, default=None)
+    p.add_argument("--output", default="results_full.jsonl")
+    p.add_argument("--dry-run", action="store_true", help="only show design size")
+    args = p.parse_args()
+
+    configs = generate_configs()
+    log.info(
+        f"design: {len(configs)} runs | factors: {[f.name for f in FACTORS]} | levels: {[len(f.levels) for f in FACTORS]}"
+    )
+
+    if not args.dry_run:
+        run_study(args.workers, args.output)

engine/studies/local_comparison.py

@@ -0,0 +1,136 @@
+import sys
+import numpy as np
+import pandas as pd
+from pathlib import Path
+import matplotlib.pyplot as plt
+
+from gymnasium.wrappers import FlattenObservation
+from stable_baselines3 import PPO
+
+# Add parent directory to path to allow importing engine
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+
+from engine.wrapper import PHANTOM
+from engine.lib.wrappers import EconomicMetricsWrapper
+from engine.lib.providers import (
+    ProviderBenchmark,
+    BenchmarkConfig,
+    RandomBaseline,
+    SurgeBaseline,
+)
+
+
+def env_factory(alpha: float):
+    """Creates a wrapped PHANTOM environment for testing at a specific alpha level."""
+    # Action levels=9 matches the trained PPO model
+    # n_products=8 matches the pretrained model's expectation of Box(16,)
+    env = PHANTOM(
+        n_products=8,
+        alpha=alpha,
+        N=100,
+        action_levels=9,
+        action_scale_low=0.8,
+        action_scale_high=1.2,
+        max_steps=20,  # Short episodes so simulation goes fast
+        robust_points=1,  # disable expensive adversarial lookaheads
+        render_mode=None,
+    )
+    env = EconomicMetricsWrapper(env)
+    return FlattenObservation(env)
+
+
+def main():
+    print("Loading pre-trained Robust RL model...")
+    model_path = Path(__file__).parent.parent / "models" / "phantom_ppo.zip"
+    if not model_path.exists():
+        print(f"Error: Model not found at {model_path}")
+        print("Please ensure you have a trained model before running this script.")
+        return
+
+    rl_model = PPO.load(model_path)
+
+    # The action space is Discrete(9). Index 4 is the middle (1.0 scale).
+    n_actions = 9
+    mid_action = n_actions // 2
+
+    providers = {
+        "Static (Base)": lambda obs: mid_action,
+        "Random": RandomBaseline(n_actions),
+        "Heuristic Surge": SurgeBaseline(
+            n_actions, high_threshold=60.0, low_threshold=30.0
+        ),
+        "Robust RL (PPO)": lambda obs: rl_model.predict(obs, deterministic=True)[0],
+    }
+
+    config = BenchmarkConfig(
+        n_episodes=10,  # Lower episodes to run faster
+        alpha_range=[0.0, 0.5, 1.0],  # Fewer alpha levels
+        baseline_name="Static (Base)",
+    )
+
+    print(f"\nStarting benchmark across alpha levels: {config.alpha_range}")
+    print(
+        f"Testing {len(providers)} strategies for {config.n_episodes} episodes each...\n"
+    )
+
+    benchmark = ProviderBenchmark(env_factory, providers, config)
+    results = benchmark.run()
+
+    # 1. Print tabular results
+    df = benchmark.to_dataframe()
+    summary = benchmark.summary_table()
+    print("\n--- Benchmark Summary Table ---")
+    print(summary)
+
+    # 2. Save results to CSV for thesis inclusion
+    out_dir = Path(__file__).parent / "results"
+    out_dir.mkdir(exist_ok=True)
+    csv_path = out_dir / "provider_comparison.csv"
+    df.to_csv(csv_path, index=False)
+    print(f"\nSaved raw results to {csv_path}")
+
+    # 3. Plot the degradation of COI / Revenue as alpha increases
+    plt.figure(figsize=(12, 5))
+
+    # Plot 1: Revenue vs Alpha
+    plt.subplot(1, 2, 1)
+    for name in providers.keys():
+        provider_data = df[df["name"] == name]
+        plt.plot(
+            provider_data["alpha"],
+            provider_data["mean_revenue"],
+            marker="o",
+            label=name,
+            linewidth=2,
+        )
+    plt.title("Revenue under Agent Contamination")
+    plt.xlabel("Contamination Level (α)")
+    plt.ylabel("Mean Episode Revenue ($)")
+    plt.grid(True, linestyle="--", alpha=0.7)
+    plt.legend()
+
+    # Plot 2: COI Preservation vs Alpha
+    plt.subplot(1, 2, 2)
+    for name in providers.keys():
+        provider_data = df[df["name"] == name]
+        plt.plot(
+            provider_data["alpha"],
+            provider_data["coi_preserved_pct"],
+            marker="s",
+            label=name,
+            linewidth=2,
+        )
+    plt.title("Cost of Information (COI) Preservation")
+    plt.xlabel("Contamination Level (α)")
+    plt.ylabel("COI Preserved (%)")
+    plt.grid(True, linestyle="--", alpha=0.7)
+    plt.legend()
+
+    plt.tight_layout()
+    plot_path = out_dir / "alpha_degradation_plot.png"
+    plt.savefig(plot_path, dpi=300)
+    print(f"Saved visualization to {plot_path}")
+
+
+if __name__ == "__main__":
+    main()

engine/studies/margin_erosion_alpha.py

@@ -0,0 +1,133 @@
+"""validate core thesis problem: margin erosion under agent contamination
+trains standard RL (no robust components) across α levels to demonstrate systematic failure
+"""
+
+from __future__ import annotations
+import json, sys, time
+from pathlib import Path
+import numpy as np
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+from engine.spec import TrainSpec
+from engine.orchestrators import run_train_once
+
+
+def _run_baseline(alpha: float, algo: str, seed: int, steps: int) -> dict:
+    spec = TrainSpec.from_flat(
+        {
+            "algo": algo,
+            "seed": seed,
+            "alpha": alpha,
+            "total_timesteps": steps,
+            "lambda_coi": 0.0,
+            "robust_radius": 0.0,
+            "robust_points": 1,
+            "robust_rollouts": 1,
+            "no_robust": True,
+            "arch": "small",
+            "n_products": 10,
+            "N": 100,
+            "max_steps": 50,
+            "eval_freq": 5000,
+            "eval_episodes": 10,
+            "log_freq": 500,
+            "robust_eval_enabled": False,
+            "agent_mu": 12.0,
+            "agent_std": 2.0,
+        }
+    )
+    result = run_train_once(
+        spec,
+        project="phantom-margin-erosion",
+        offline=True,
+        no_wandb=True,
+        kind="study",
+        scenario=f"alpha{int(alpha * 100):02d}",
+        group=f"baseline_{algo}",
+        extra_tags=("margin_erosion", "baseline"),
+    )
+    return {
+        "alpha": alpha,
+        "algo": algo,
+        "seed": seed,
+        "eval_reward": result.get("eval/reward_mean", np.nan),
+        "eval_revenue": result.get("eval/revenue_mean", np.nan),
+        "eval_coi_level": result.get("eval/coi_level_mean", np.nan),
+        "eval_margin": result.get("eval/margin_mean", np.nan),
+        "eval_agent_prob": result.get("eval/agent_prob_mean", np.nan),
+    }
+
+
+def run_margin_erosion_study(
+    alphas: list[float] | None = None,
+    algos: list[str] | None = None,
+    seeds: int = 3,
+    steps: int = 30_000,
+) -> dict:
+    alphas = alphas or [0.1, 0.3, 0.5, 0.7, 0.9]
+    algos = algos or ["ppo", "dqn", "qtable"]
+    output_dir = Path(__file__).parent / "results"
+    output_dir.mkdir(exist_ok=True)
+    ts = time.strftime("%Y%m%d_%H%M%S")
+
+    results = []
+    for α in alphas:
+        for algo in algos:
+            for si in range(seeds):
+                seed = 42 + si
+                print(f"α={α:.1f} {algo} seed={seed}")
+                m = _run_baseline(α, algo, seed, steps)
+                results.append(m)
+                print(
+                    f"  margin={m['eval_margin']:.3f} rev={m['eval_revenue']:.0f} coi={m['eval_coi_level']:.1f}"
+                )
+
+    summary = {}
+    for α in alphas:
+        runs = [r for r in results if abs(r["alpha"] - α) < 0.01]
+        if not runs:
+            continue
+        s = {}
+        for metric in ["margin", "revenue", "coi_level", "agent_prob"]:
+            vals = [r[f"eval_{metric}"] for r in runs]
+            s[f"{metric}_mean"] = float(np.mean(vals))
+            s[f"{metric}_std"] = float(np.std(vals))
+        s["n_runs"] = len(runs)
+        summary[f"alpha_{α:.1f}"] = s
+
+    output = {
+        "timestamp": ts,
+        "config": {"alphas": alphas, "algos": algos, "seeds": seeds, "steps": steps},
+        "results": results,
+        "summary": summary,
+    }
+
+    path = output_dir / f"margin_erosion_alpha_{ts}.json"
+    with open(path, "w") as f:
+        json.dump(output, f, indent=2)
+
+    print(f"\n→ {path}")
+    for α in alphas:
+        k = f"alpha_{α:.1f}"
+        if k in summary:
+            s = summary[k]
+            print(
+                f"  {k}: margin={s['margin_mean']:.3f}±{s['margin_std']:.3f} "
+                f"coi={s['coi_level_mean']:.1f}±{s['coi_level_std']:.1f}"
+            )
+    return output
+
+
+if __name__ == "__main__":
+    import argparse
+
+    p = argparse.ArgumentParser(description="margin erosion vs α")
+    p.add_argument("--quick", action="store_true", help="fast test")
+    args = p.parse_args()
+
+    run_margin_erosion_study(
+        alphas=[0.1, 0.7] if args.quick else [0.1, 0.3, 0.5, 0.7, 0.9],
+        algos=["qtable"] if args.quick else ["ppo", "dqn", "qtable"],
+        seeds=1 if args.quick else 3,
+        steps=5_000 if args.quick else 30_000,
+    )

engine/studies/mixed_lh.py

@@ -0,0 +1,127 @@
+"""mixed design: full factorial on primary factors, latin hypercube on secondary"""
+
+import sys
+
+sys.path.insert(0, "..")
+import logging
+from itertools import product
+import json
+import hashlib
+from pathlib import Path
+from concurrent.futures import ProcessPoolExecutor
+import numpy as np
+from scipy.stats.qmc import LatinHypercube
+from factors import FACTORS, DEMAND_FUNCTIONS, SEEDS_PER_CONFIG
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+log = logging.getLogger(__name__)
+
+LH_SAMPLES = 10
+
+
+def generate_configs(lh_samples: int = LH_SAMPLES):
+    primary = [f for f in FACTORS if f.primary]
+    secondary = [f for f in FACTORS if not f.primary]
+
+    primary_grid = list(product(*[f.levels for f in primary]))
+    lhs = LatinHypercube(d=len(secondary), seed=42)
+
+    configs = []
+    for p_combo in primary_grid:
+        samples = lhs.random(n=lh_samples)
+        for s in samples:
+            sec_vals = {
+                secondary[i].name: secondary[i].levels[
+                    int(s[i] * len(secondary[i].levels))
+                ]
+                for i in range(len(secondary))
+            }
+            base = {primary[i].name: p_combo[i] for i in range(len(primary))}
+            base.update(sec_vals)
+
+            for seed in range(SEEDS_PER_CONFIG):
+                cfg = {**base, "seed": seed}
+                cfg["id"] = hashlib.md5(
+                    json.dumps(cfg, sort_keys=True).encode()
+                ).hexdigest()[:8]
+                configs.append(cfg)
+    return configs
+
+
+def run_single(cfg: dict) -> dict:
+    from engine.wrapper import PHANTOM
+    import numpy as np
+
+    np.random.seed(cfg["seed"])
+    demand_fn = DEMAND_FUNCTIONS[cfg["demand_fn"]]
+
+    env = PHANTOM(
+        n_products=cfg["n_products"],
+        alpha=cfg["alpha"],
+        N=cfg["N"],
+    )
+    env.market.demand = (demand_fn, (cfg["demand_mu"], cfg["demand_sigma"]))
+
+    obs, _ = env.reset()
+    total_reward, steps = 0.0, 0
+
+    for _ in range(100):
+        action = env.action_space.sample()
+        obs, reward, term, trunc, _ = env.step(action)
+        total_reward += reward
+        steps += 1
+        if term:
+            break
+
+    env.close()
+    return {
+        "id": cfg["id"],
+        "config": cfg,
+        "total_reward": total_reward,
+        "avg_reward": total_reward / steps,
+        "steps": steps,
+    }
+
+
+def run_study(
+    max_workers: int = None,
+    output: str = "results_mixed.jsonl",
+    lh_samples: int = LH_SAMPLES,
+):
+    configs = generate_configs(lh_samples)
+    n_primary_cells = int(np.prod([len(f.levels) for f in FACTORS if f.primary]))
+    log.info(
+        f"mixed LH: {len(configs)} configs ({n_primary_cells} primary × {lh_samples} LH × {SEEDS_PER_CONFIG} seeds)"
+    )
+
+    results = []
+    with ProcessPoolExecutor(max_workers=max_workers) as ex:
+        for i, result in enumerate(ex.map(run_single, configs)):
+            results.append(result)
+            if (i + 1) % 100 == 0:
+                log.info(f"progress: {i + 1}/{len(configs)}")
+
+    Path(output).write_text("\n".join(json.dumps(r) for r in results))
+    log.info(f"wrote {len(results)} results to {output}")
+    return results
+
+
+if __name__ == "__main__":
+    import argparse
+
+    p = argparse.ArgumentParser()
+    p.add_argument("--workers", type=int, default=None)
+    p.add_argument("--output", default="results_mixed.jsonl")
+    p.add_argument("--lh-samples", type=int, default=10)
+    p.add_argument("--dry-run", action="store_true", help="only show design size")
+    args = p.parse_args()
+
+    primary = [f for f in FACTORS if f.primary]
+    secondary = [f for f in FACTORS if not f.primary]
+    configs = generate_configs(args.lh_samples)
+    log.info(
+        f"design: {len(configs)} runs | primary: {[f.name for f in primary]} | secondary (LH): {[f.name for f in secondary]}"
+    )
+
+    if not args.dry_run:
+        run_study(args.workers, args.output, args.lh_samples)

engine/sweeps/final_thesis_proof.yaml

@@ -0,0 +1,60 @@
+method: grid
+metric:
+  name: eval/stress_reward_worst
+  goal: maximize
+command:
+  - ${env}
+  - python
+  - -m
+  - engine.train
+parameters:
+  algo:
+    value: ppo
+  backend:
+    value: sb3
+  device:
+    value: cpu
+  seed:
+    values: [42, 1337, 7777]
+  alpha:
+    values: [0.1, 0.2, 0.3, 0.4, 0.6, 0.8]
+  n_products:
+    values: [25, 50, 100]
+  N:
+    value: 100
+  no_robust:
+    values: [false, true]
+  lambda_coi:
+    values: [0.15, 0.30]
+  robust_radius:
+    value: 0.2
+  robust_points:
+    value: 7
+  robust_rollouts:
+    value: 1
+  eta_ux:
+    value: 0.5
+  reward_profit_weight:
+    value: 1.0
+  action_levels:
+    value: 9
+  action_scale_low:
+    value: 0.8
+  action_scale_high:
+    value: 1.2
+  total_timesteps:
+    value: 100000
+  eval_episodes:
+    value: 12
+  eval_freq:
+    value: 1000
+  log_freq:
+    value: 100
+  hist_freq:
+    value: 500
+  learning_rate:
+    value: 0.0003
+  batch_size:
+    value: 256
+  n_steps:
+    value: 2048

engine/sweeps/model_mix.yaml

@@ -0,0 +1,84 @@
+method: random
+metric:
+  name: objective/score
+  goal: maximize
+command:
+  - ${env}
+  - python
+  - -m
+  - engine.train
+parameters:
+  algo:
+    values: [ppo, a2c, dqn, qtable]
+  total_timesteps:
+    values: [30000, 50000, 80000]
+  seed:
+    values: [13, 42, 77]
+  n_products:
+    values: [8, 10, 12]
+  alpha:
+    distribution: uniform
+    min: 0.1
+    max: 0.6
+  lambda_coi:
+    distribution: uniform
+    min: 0.05
+    max: 0.6
+  robust_radius:
+    distribution: uniform
+    min: 0.0
+    max: 0.3
+  robust_points:
+    values: [3, 5, 7]
+  info_value:
+    distribution: uniform
+    min: 0.5
+    max: 2.0
+  revenue_weight:
+    values: [0.005, 0.01, 0.02]
+  learning_rate:
+    distribution: log_uniform_values
+    min: 1.0e-5
+    max: 1.0e-3
+  gamma:
+    values: [0.97, 0.99, 0.995]
+  buffer_size:
+    values: [20000, 50000, 100000]
+  batch_size:
+    values: [128, 256, 512]
+  tau:
+    values: [0.002, 0.005, 0.01]
+  train_freq:
+    values: [1, 4, 8]
+  learning_starts:
+    values: [500, 1000, 3000]
+  n_steps:
+    values: [512, 1024, 2048]
+  n_epochs:
+    values: [5, 10, 20]
+  gae_lambda:
+    values: [0.9, 0.95, 0.98]
+  clip_range:
+    values: [0.1, 0.2, 0.3]
+  ent_coef:
+    values: [0.0, 0.005, 0.01]
+  target_update_interval:
+    values: [500, 1000, 2000]
+  exploration_fraction:
+    values: [0.1, 0.2, 0.3]
+  exploration_final_eps:
+    values: [0.01, 0.03, 0.05]
+  action_levels:
+    values: [7, 9, 11]
+  action_scale_low:
+    values: [0.75, 0.8, 0.85]
+  action_scale_high:
+    values: [1.15, 1.2, 1.25]
+  q_lr:
+    values: [0.03, 0.05, 0.1, 0.2]
+  eps_start:
+    value: 1.0
+  eps_end:
+    values: [0.02, 0.05, 0.1]
+  eps_decay:
+    values: [0.999, 0.9995, 0.9999]

engine/sweeps/models_only.yaml

@@ -0,0 +1,85 @@
+method: grid
+metric:
+  name: objective/score
+  goal: maximize
+run_cap: 4
+command:
+  - ${env}
+  - python
+  - -m
+  - engine.train
+parameters:
+  algo:
+    values: [ppo, a2c, dqn, qtable]
+  seed:
+    value: 42
+  total_timesteps:
+    value: 12000
+  eval_episodes:
+    value: 3
+  eval_freq:
+    value: 500
+  log_freq:
+    value: 100
+  revenue_weight:
+    value: 0.01
+  n_products:
+    value: 8
+  N:
+    value: 80
+  alpha:
+    value: 0.3
+  lambda_coi:
+    value: 0.2
+  robust_radius:
+    value: 0.0
+  robust_points:
+    value: 1
+  info_value:
+    value: 1.0
+  learning_rate:
+    value: 0.0003
+  gamma:
+    value: 0.99
+  buffer_size:
+    value: 20000
+  batch_size:
+    value: 128
+  tau:
+    value: 0.005
+  train_freq:
+    value: 1
+  learning_starts:
+    value: 500
+  n_steps:
+    value: 512
+  n_epochs:
+    value: 10
+  gae_lambda:
+    value: 0.95
+  clip_range:
+    value: 0.2
+  ent_coef:
+    value: 0.0
+  target_update_interval:
+    value: 500
+  exploration_fraction:
+    value: 0.2
+  exploration_final_eps:
+    value: 0.05
+  action_levels:
+    value: 7
+  action_scale_low:
+    value: 0.9
+  action_scale_high:
+    value: 1.1
+  q_lr:
+    value: 0.1
+  q_bins:
+    value: 6
+  eps_start:
+    value: 1.0
+  eps_end:
+    value: 0.05
+  eps_decay:
+    value: 0.9995

engine/sweeps/ppo_supra_guard.yaml

@@ -0,0 +1,53 @@
+method: random
+metric:
+  name: eval/supra_share_mean
+  goal: minimize
+run_cap: 256
+command:
+  - ${env}
+  - python
+  - -m
+  - engine.train
+parameters:
+  algo:
+    value: ppo
+  seed:
+    values: [42, 1337, 7777]
+  alpha:
+    values: [0.1, 0.2, 0.3, 0.4, 0.6]
+  n_products:
+    values: [25, 50]
+  N:
+    value: 100
+  no_robust:
+    values: [false, true]
+  lambda_coi:
+    values: [0.05, 0.15, 0.3]
+  robust_radius:
+    values: [0.1, 0.2, 0.3]
+  robust_points:
+    value: 7
+  robust_rollouts:
+    value: 1
+  eta_ux:
+    values: [0.05, 0.15, 0.3, 0.5, 0.75]
+  reward_profit_weight:
+    value: 1.0
+  total_timesteps:
+    value: 100000
+  eval_episodes:
+    value: 10
+  eval_freq:
+    value: 1000
+  log_freq:
+    value: 100
+  hist_freq:
+    value: 500
+  learning_rate:
+    value: 0.0003
+  batch_size:
+    value: 256
+  n_steps:
+    value: 2048
+  device:
+    value: cpu

engine/sweeps/sac_tune.yaml

@@ -0,0 +1,54 @@
+method: bayes
+metric:
+  name: objective/score
+  goal: maximize
+command:
+  - ${env}
+  - python
+  - -m
+  - engine.train
+parameters:
+  algo:
+    value: sac
+  total_timesteps:
+    values: [50000, 80000, 120000]
+  seed:
+    values: [13, 42, 77]
+  alpha:
+    distribution: uniform
+    min: 0.15
+    max: 0.55
+  n_products:
+    values: [8, 10, 12]
+  lambda_coi:
+    distribution: uniform
+    min: 0.05
+    max: 0.5
+  robust_radius:
+    distribution: uniform
+    min: 0.05
+    max: 0.3
+  robust_points:
+    values: [3, 5, 7]
+  info_value:
+    distribution: uniform
+    min: 0.5
+    max: 2.0
+  revenue_weight:
+    values: [0.005, 0.01, 0.02]
+  learning_rate:
+    distribution: log_uniform_values
+    min: 3.0e-5
+    max: 1.0e-3
+  gamma:
+    values: [0.98, 0.99, 0.995]
+  buffer_size:
+    values: [50000, 100000, 200000]
+  batch_size:
+    values: [128, 256, 512]
+  tau:
+    values: [0.002, 0.005, 0.01]
+  train_freq:
+    values: [1, 4, 8]
+  learning_starts:
+    values: [1000, 3000, 5000]

engine/sweeps/small_arch_compare.yaml

@@ -0,0 +1,86 @@
+method: random
+metric:
+  name: objective/score
+  goal: maximize
+command:
+  - ${env}
+  - python
+  - -m
+  - engine.train
+parameters:
+  algo:
+    values: [ppo, a2c, dqn, qtable]
+  arch:
+    values: [tiny, small, medium]
+  activation:
+    values: [relu, tanh]
+  total_timesteps:
+    values: [8000, 12000, 20000]
+  seed:
+    values: [13, 42, 77]
+  n_products:
+    values: [6, 8, 10]
+  alpha:
+    distribution: uniform
+    min: 0.1
+    max: 0.5
+  lambda_coi:
+    distribution: uniform
+    min: 0.05
+    max: 0.4
+  robust_radius:
+    values: [0.0, 0.1, 0.2]
+  robust_points:
+    values: [3, 5]
+  info_value:
+    values: [0.75, 1.0, 1.5]
+  revenue_weight:
+    values: [0.005, 0.01, 0.02]
+  learning_rate:
+    distribution: log_uniform_values
+    min: 1.0e-5
+    max: 5.0e-4
+  gamma:
+    values: [0.98, 0.99]
+  buffer_size:
+    values: [10000, 30000, 50000]
+  batch_size:
+    values: [64, 128, 256]
+  tau:
+    values: [0.002, 0.005, 0.01]
+  train_freq:
+    values: [1, 4]
+  learning_starts:
+    values: [500, 1000, 2000]
+  n_steps:
+    values: [256, 512, 1024]
+  n_epochs:
+    values: [5, 10]
+  gae_lambda:
+    values: [0.9, 0.95]
+  clip_range:
+    values: [0.1, 0.2]
+  ent_coef:
+    values: [0.0, 0.005]
+  target_update_interval:
+    values: [500, 1000]
+  exploration_fraction:
+    values: [0.1, 0.2]
+  exploration_final_eps:
+    values: [0.02, 0.05]
+  action_levels:
+    values: [5, 7, 9]
+  action_scale_low:
+    values: [0.85, 0.9]
+  action_scale_high:
+    values: [1.1, 1.15]
+  q_lr:
+    values: [0.05, 0.1, 0.2]
+  q_bins:
+    values: [4, 6, 8]
+  eps_start:
+    value: 1.0
+  eps_end:
+    values: [0.02, 0.05]
+  eps_decay:
+    values: [0.999, 0.9995]

engine/telemetry/__init__.py

@@ -0,0 +1,23 @@
+from .metrics import canonicalize_metrics
+from .wandb import (
+    current_config,
+    finish_run,
+    get_wandb_module,
+    init_run,
+    log_metrics,
+    run_agent,
+    update_run_config,
+    update_summary,
+)
+
+__all__ = [
+    "canonicalize_metrics",
+    "current_config",
+    "finish_run",
+    "get_wandb_module",
+    "init_run",
+    "log_metrics",
+    "run_agent",
+    "update_run_config",
+    "update_summary",
+]

engine/telemetry/metrics.py

@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+from typing import Any, Mapping
+
+from ..spec import TrainSpec
+
+
+_ALIASES = {
+    "train/reward": "train/reward_mean",
+    "train/revenue": "train/revenue_mean",
+    "train/dqn_loss": "train/loss",
+    "eval/reward": "eval/reward_mean",
+    "eval/revenue": "eval/revenue_mean",
+    "train/steps_per_second": "runtime/steps_per_second",
+}
+
+
+def _as_float(value: Any, default: float | None = None) -> float | None:
+    if value is None:
+        return default
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return default
+
+
+def canonicalize_metrics(raw: Mapping[str, Any], spec: TrainSpec) -> dict[str, Any]:
+    metrics: dict[str, Any] = {}
+    for key, value in raw.items():
+        canonical = _ALIASES.get(str(key), str(key))
+        if canonical in metrics and canonical != key:
+            continue
+        metrics[canonical] = value
+
+    metrics.setdefault("train/global_step", spec.runtime.total_timesteps)
+
+    eval_reward = (
+        _as_float(
+            metrics.get(
+                "eval/stress_reward_worst",
+                metrics.get(
+                    "eval/robust_reward_worst", metrics.get("eval/reward_mean")
+                ),
+            ),
+            0.0,
+        )
+        or 0.0
+    )
+    metrics["objective/score"] = eval_reward
+
+    margin_mean = _as_float(metrics.get("eval/margin_mean"), None)
+    if margin_mean is not None:
+        metrics["objective/constraint_margin"] = margin_mean - spec.env.margin_floor
+
+    coi_level = _as_float(metrics.get("eval/coi_level_mean"), None)
+    metrics["objective/coi_preserved"] = 0.0 if coi_level is None else coi_level
+
+    metrics["study/alpha"] = spec.study.alpha
+    metrics["study/mode"] = "baseline" if bool(spec.study.no_robust) else "defended"
+    metrics["study/baseline_mode"] = float(bool(spec.study.no_robust))
+    metrics["study/lambda_coi"] = spec.study.lambda_coi
+    metrics["study/ambiguity_radius"] = spec.study.robust_radius
+    metrics["study/info_value"] = spec.study.info_value
+    metrics["tiers"] = spec.algorithm.name
+
+    metrics["runtime/backend"] = spec.runtime.backend
+    metrics["runtime/device"] = spec.runtime.device
+    metrics["runtime/seed"] = spec.runtime.seed
+
+    return metrics

engine/telemetry/wandb.py

@@ -0,0 +1,202 @@
+from __future__ import annotations
+
+import os
+import time
+from typing import Any, Callable, Iterable, Mapping
+
+
+def get_wandb_module():
+    try:
+        import wandb
+
+        return wandb
+    except ImportError:
+        return None
+
+
+def _require_wandb():
+    wandb = get_wandb_module()
+    if wandb is None:
+        raise ImportError("wandb is required for this workflow")
+    return wandb
+
+
+def _warn(message: str) -> None:
+    print(f"PHANTOM_WANDB_WARNING: {message}")
+
+
+def _sanitize_key(raw_key: str) -> str | None:
+    key = str(raw_key)
+    replacements = {
+        "no_robust": "baseline_mode",
+        "study/no_robust": "study/baseline_mode",
+        "study/robust_radius": "study/ambiguity_radius",
+        "robust_radius": "ambiguity_radius",
+        "robust_points": "ambiguity_points",
+        "robust_rollouts": "ambiguity_rollouts",
+        "robust_eval_enabled": "stress_eval_enabled",
+        "eval/robust_alpha_high": "eval/stress_alpha_high",
+        "eval/robust_alpha_low": "eval/stress_alpha_low",
+        "eval/robust_reward_worst": "eval/stress_reward_worst",
+        "eval/robust_revenue_worst": "eval/stress_revenue_worst",
+        "eval/robust_coi_leakage_worst": "eval/stress_coi_leakage_worst",
+    }
+    key = replacements.get(key, key)
+    if "robust" in key.lower():
+        return None
+    return key
+
+
+def _sanitize_payload(payload: Mapping[str, Any]) -> dict[str, Any]:
+    sanitized: dict[str, Any] = {}
+    for key, value in payload.items():
+        clean_key = _sanitize_key(str(key))
+        if clean_key is None:
+            continue
+        sanitized[clean_key] = value
+    return sanitized
+
+
+def init_run(
+    *,
+    mode: str,
+    project: str | None = None,
+    config: Mapping[str, Any] | None = None,
+    name: str | None = None,
+    tags: Iterable[str] | None = None,
+    group: str | None = None,
+    sweep_mode: bool = False,
+):
+    wandb = _require_wandb()
+    kwargs: dict[str, Any] = {"mode": mode}
+    if group:
+        kwargs["group"] = group
+    if sweep_mode:
+        try:
+            run = wandb.init(**kwargs)
+        except Exception as exc:
+            _warn(f"init failed in sweep mode ({exc})")
+            return None
+        if name and run is not None:
+            run.name = name
+        return run
+
+    init_kwargs = dict(kwargs)
+    init_kwargs["project"] = project
+    if config is not None:
+        init_kwargs["config"] = _sanitize_payload(dict(config))
+    if name:
+        init_kwargs["name"] = name
+    if tags:
+        init_kwargs["tags"] = list(tags)
+    try:
+        return wandb.init(**init_kwargs)
+    except Exception as exc:
+        _warn(f"init failed ({exc})")
+        return None
+
+
+def finish_run() -> None:
+    wandb = get_wandb_module()
+    if wandb is not None and wandb.run is not None:
+        try:
+            wandb.finish()
+        except Exception as exc:
+            _warn(f"finish failed ({exc})")
+
+
+def current_config() -> dict[str, Any]:
+    wandb = get_wandb_module()
+    if wandb is None or wandb.run is None:
+        return {}
+    return {key: wandb.config[key] for key in wandb.config.keys()}
+
+
+def update_run_config(config: Mapping[str, Any]) -> None:
+    wandb = get_wandb_module()
+    if wandb is None or wandb.run is None:
+        return
+    payload = _sanitize_payload(dict(config))
+    if not payload:
+        return
+    try:
+        wandb.config.update(payload, allow_val_change=True)
+    except TypeError:
+        try:
+            wandb.config.update(payload)
+        except Exception as exc:
+            _warn(f"config update failed ({exc})")
+    except Exception as exc:
+        _warn(f"config update failed ({exc})")
+
+
+def log_metrics(metrics: Mapping[str, Any], *, step: int) -> None:
+    wandb = get_wandb_module()
+    if wandb is None or wandb.run is None:
+        return
+    payload = _sanitize_payload(dict(metrics))
+    if not payload:
+        return
+    try:
+        wandb.log(payload, step=step)
+    except Exception as exc:
+        _warn(f"log failed at step {step} ({exc})")
+
+
+def update_summary(metrics: Mapping[str, Any]) -> None:
+    wandb = get_wandb_module()
+    if wandb is None or wandb.run is None:
+        return
+    payload = _sanitize_payload(dict(metrics))
+    if not payload:
+        return
+    try:
+        for key, value in payload.items():
+            wandb.run.summary[key] = value
+    except Exception as exc:
+        _warn(f"summary update failed ({exc})")
+
+
+def run_agent(
+    sweep_id: str,
+    fn: Callable[[], None],
+    *,
+    count: int | None = None,
+) -> None:
+    wandb = _require_wandb()
+    retry_max = max(0, int(os.getenv("PHANTOM_WANDB_AGENT_RETRIES", "8")))
+    retry_delay = max(1.0, float(os.getenv("PHANTOM_WANDB_AGENT_RETRY_DELAY", "5")))
+    retry_backoff = max(
+        1.0, float(os.getenv("PHANTOM_WANDB_AGENT_RETRY_BACKOFF", "1.5"))
+    )
+    retry_max_delay = max(
+        retry_delay,
+        float(os.getenv("PHANTOM_WANDB_AGENT_MAX_RETRY_DELAY", "60")),
+    )
+
+    target = None if count is None else max(0, int(count))
+    completed = 0
+
+    def _wrapped() -> None:
+        nonlocal completed
+        fn()
+        completed += 1
+
+    attempt = 0
+    while True:
+        remaining = None if target is None else max(0, int(target - completed))
+        if target is not None and remaining == 0:
+            return
+        try:
+            wandb.agent(sweep_id, function=_wrapped, count=remaining)
+            return
+        except Exception as exc:
+            attempt += 1
+            if attempt > retry_max:
+                raise
+            wait = min(retry_max_delay, retry_delay * (retry_backoff ** (attempt - 1)))
+            _warn(
+                f"agent disconnected (attempt {attempt}/{retry_max}, "
+                f"completed={completed}, remaining={remaining}): {exc}"
+            )
+            time.sleep(wait)

engine/train.py

@@ -0,0 +1,251 @@
+from __future__ import annotations
+
+import argparse
+from typing import Any
+
+from .logging_utils import configure_logging
+from .orchestrators import run_benchmark_cli, run_sweep_agent, run_train_once
+from .spec import TrainSpec
+
+
+def _parse_tags(raw: str | None) -> list[str]:
+    if raw is None:
+        return []
+    return [piece.strip() for piece in str(raw).split(",") if piece.strip()]
+
+
+def _probe_run_kind(argv: list[str]) -> str:
+    probe = argparse.ArgumentParser(add_help=False)
+    probe.add_argument("--run-kind", choices=["train", "benchmark"])
+    probe.add_argument("--run-mode", choices=["train", "benchmark"])
+    args, _ = probe.parse_known_args(argv)
+    return str(args.run_kind or args.run_mode or "train")
+
+
+def _strip_run_kind(argv: list[str]) -> list[str]:
+    stripped: list[str] = []
+    skip_next = False
+    for item in argv:
+        if skip_next:
+            skip_next = False
+            continue
+        if item in {"--run-kind", "--run-mode"}:
+            skip_next = True
+            continue
+        if item.startswith("--run-kind=") or item.startswith("--run-mode="):
+            continue
+        stripped.append(item)
+    return stripped
+
+
+def _build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description="PHANTOM unified training entrypoint")
+    parser.add_argument("--run-kind", choices=["train", "benchmark"], default="train")
+    parser.add_argument("--run-mode", choices=["train", "benchmark"])
+
+    parser.add_argument("--project", default="capstone")
+    parser.add_argument("--scenario", default="default")
+    parser.add_argument("--group", type=str)
+    parser.add_argument("--tags", type=str)
+
+    parser.add_argument("--backend", choices=["auto", "sb3"], default="auto")
+    parser.add_argument("--algo", choices=["ppo", "a2c", "dqn", "qtable", "sac"])
+    parser.add_argument("--seed", type=int)
+    parser.add_argument("--total-timesteps", type=int)
+    parser.add_argument("--model-dir", type=str)
+    parser.add_argument("--log-freq", type=int)
+    parser.add_argument("--hist-freq", type=int)
+    parser.add_argument("--checkpoint-interval", type=int)
+    parser.add_argument("--device", type=str)
+
+    parser.add_argument("--alpha", type=float)
+    parser.add_argument("--N", type=int)
+    parser.add_argument("--n-products", type=int)
+    parser.add_argument("--lambda-coi", type=float)
+    parser.add_argument("--info-value", type=float)
+    parser.add_argument("--robust-radius", type=float)
+    parser.add_argument("--robust-points", type=int)
+    parser.add_argument("--robust-rollouts", type=int)
+    parser.add_argument("--no-robust", action="store_true")
+    parser.add_argument("--eta-ux", type=float)
+    parser.add_argument("--reward-profit-weight", type=float)
+
+    parser.add_argument("--price-low", type=float)
+    parser.add_argument("--price-high", type=float)
+    parser.add_argument("--action-levels", type=int)
+    parser.add_argument("--action-scale-low", type=float)
+    parser.add_argument("--action-scale-high", type=float)
+    parser.add_argument("--max-steps", type=int)
+    parser.add_argument("--margin-floor", type=float)
+    parser.add_argument("--margin-floor-patience", type=int)
+
+    parser.add_argument("--learning-rate", type=float)
+    parser.add_argument("--gamma", type=float)
+    parser.add_argument("--buffer-size", type=int)
+    parser.add_argument("--batch-size", type=int)
+    parser.add_argument("--tau", type=float)
+    parser.add_argument("--train-freq", type=int)
+    parser.add_argument("--learning-starts", type=int)
+    parser.add_argument("--target-update-interval", type=int)
+    parser.add_argument("--exploration-fraction", type=float)
+    parser.add_argument("--exploration-final-eps", type=float)
+    parser.add_argument("--n-steps", type=int)
+    parser.add_argument("--n-epochs", type=int)
+    parser.add_argument("--gae-lambda", type=float)
+    parser.add_argument("--clip-range", type=float)
+    parser.add_argument("--ent-coef", type=float)
+    parser.add_argument("--q-lr", type=float)
+    parser.add_argument("--q-bins", type=int)
+    parser.add_argument("--eps-start", type=float)
+    parser.add_argument("--eps-end", type=float)
+    parser.add_argument("--eps-decay", type=float)
+    parser.add_argument("--arch", type=str)
+    parser.add_argument("--activation", type=str)
+    parser.add_argument("--vf-coef", type=float)
+    parser.add_argument("--max-grad-norm", type=float)
+
+    parser.add_argument("--eval-freq", type=int)
+    parser.add_argument("--eval-episodes", type=int)
+
+    parser.add_argument("--sweep-agent", action="store_true")
+    parser.add_argument("--sweep-id", type=str)
+    parser.add_argument("--count", type=int, default=0)
+    parser.add_argument("--offline", action="store_true")
+    parser.add_argument("--no-wandb", action="store_true")
+    return parser
+
+
+def _overrides_from_args(args: argparse.Namespace) -> dict[str, Any]:
+    backend = None if args.backend == "auto" else args.backend
+
+    overrides = {
+        "project": args.project,
+        "backend": backend,
+        "algo": args.algo,
+        "seed": args.seed,
+        "total_timesteps": args.total_timesteps,
+        "model_dir": args.model_dir,
+        "log_freq": args.log_freq,
+        "hist_freq": args.hist_freq,
+        "checkpoint_interval": args.checkpoint_interval,
+        "device": args.device,
+        "alpha": args.alpha,
+        "N": args.N,
+        "n_products": args.n_products,
+        "lambda_coi": args.lambda_coi,
+        "info_value": args.info_value,
+        "robust_radius": args.robust_radius,
+        "robust_points": args.robust_points,
+        "robust_rollouts": args.robust_rollouts,
+        "no_robust": args.no_robust,
+        "eta_ux": args.eta_ux,
+        "reward_profit_weight": args.reward_profit_weight,
+        "price_low": args.price_low,
+        "price_high": args.price_high,
+        "action_levels": args.action_levels,
+        "action_scale_low": args.action_scale_low,
+        "action_scale_high": args.action_scale_high,
+        "max_steps": args.max_steps,
+        "margin_floor": args.margin_floor,
+        "margin_floor_patience": args.margin_floor_patience,
+        "learning_rate": args.learning_rate,
+        "gamma": args.gamma,
+        "buffer_size": args.buffer_size,
+        "batch_size": args.batch_size,
+        "tau": args.tau,
+        "train_freq": args.train_freq,
+        "learning_starts": args.learning_starts,
+        "target_update_interval": args.target_update_interval,
+        "exploration_fraction": args.exploration_fraction,
+        "exploration_final_eps": args.exploration_final_eps,
+        "n_steps": args.n_steps,
+        "n_epochs": args.n_epochs,
+        "gae_lambda": args.gae_lambda,
+        "clip_range": args.clip_range,
+        "ent_coef": args.ent_coef,
+        "q_lr": args.q_lr,
+        "q_bins": args.q_bins,
+        "eps_start": args.eps_start,
+        "eps_end": args.eps_end,
+        "eps_decay": args.eps_decay,
+        "arch": args.arch,
+        "activation": args.activation,
+        "vf_coef": args.vf_coef,
+        "max_grad_norm": args.max_grad_norm,
+        "eval_freq": args.eval_freq,
+        "eval_episodes": args.eval_episodes,
+    }
+    return {key: value for key, value in overrides.items() if value is not None}
+
+
+def main(argv: list[str] | None = None) -> None:
+    import subprocess
+    import sys
+
+    # Ensure data is downloaded
+    from pathlib import Path
+
+    project_root = Path(__file__).parents[1]
+    data_dir = project_root / "experiments" / "collected_data"
+    needs_pull = (not data_dir.exists()) or (not any(data_dir.iterdir()))
+    if needs_pull:
+        try:
+            subprocess.run(["make", "data.pull"], cwd=str(project_root), check=True)
+        except (subprocess.SubprocessError, OSError) as exc:
+            sys.path.insert(0, str(project_root))
+            try:
+                from scripts.hf_data import pull
+
+                pull()
+            except (ImportError, OSError, RuntimeError, ValueError) as fallback_exc:
+                print(
+                    f"Warning: data.pull failed ({exc}); fallback pull failed ({fallback_exc})"
+                )
+
+    configure_logging()
+    raw_args = list(sys.argv[1:] if argv is None else argv)
+    run_kind = _probe_run_kind(raw_args)
+    if run_kind == "benchmark":
+        run_benchmark_cli(_strip_run_kind(raw_args))
+        return
+
+    parser = _build_parser()
+    args, unknown = parser.parse_known_args(raw_args)
+    if unknown:
+        raise ValueError(f"Unknown arguments for training mode: {' '.join(unknown)}")
+
+    overrides = _overrides_from_args(args)
+    scenario = str(args.scenario)
+    group = args.group
+    extra_tags = tuple(_parse_tags(args.tags))
+
+    if args.sweep_agent:
+        run_sweep_agent(
+            project=args.project,
+            sweep_id=str(args.sweep_id or ""),
+            count=int(args.count),
+            offline=bool(args.offline),
+            no_wandb=bool(args.no_wandb),
+            base_overrides=overrides,
+            kind="sweep",
+            scenario=scenario,
+            group=group,
+            extra_tags=extra_tags,
+        )
+        return
+
+    spec = TrainSpec.from_flat(overrides)
+    run_train_once(
+        spec,
+        project=args.project,
+        offline=bool(args.offline),
+        no_wandb=bool(args.no_wandb),
+        kind="train",
+        scenario=scenario,
+        group=group,
+        extra_tags=extra_tags,
+    )
+
+
+if __name__ == "__main__":
+    main()

engine/train_core.py

@@ -0,0 +1,40 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any
+
+from .spec import TrainSpec
+from .telemetry.metrics import canonicalize_metrics
+
+
+@dataclass(frozen=True)
+class TrainResult:
+    spec: TrainSpec
+    metrics: dict[str, Any]
+    artifacts: dict[str, str]
+    events: list[dict[str, Any]]
+
+
+def run_train(spec: TrainSpec) -> TrainResult:
+    cfg = spec.to_flat_dict()
+    algo = spec.algorithm.name
+
+    if algo == "qtable":
+        from .backends.qtable import train_qtable
+
+        _, raw_metrics = train_qtable(cfg)
+    else:
+        from .backends.sb3 import train_sb3
+
+        _, raw_metrics = train_sb3(cfg)
+
+    events_raw = raw_metrics.pop("_train_events", [])
+    events = [evt for evt in events_raw if isinstance(evt, dict)]
+
+    metrics = canonicalize_metrics(raw_metrics, spec)
+    artifacts: dict[str, str] = {}
+    model_path = raw_metrics.get("model/path")
+    if isinstance(model_path, str):
+        artifacts["model/path"] = model_path
+
+    return TrainResult(spec=spec, metrics=metrics, artifacts=artifacts, events=events)

engine/wandb_checkpoint.py

@@ -0,0 +1,130 @@
+from __future__ import annotations
+
+import hashlib
+import json
+import re
+from pathlib import Path
+from tempfile import TemporaryDirectory
+from typing import Any, Mapping
+
+try:
+    import wandb
+    from wandb.errors import CommError
+
+    HAS_WANDB = True
+except ImportError:
+    HAS_WANDB = False
+    wandb = None  # type: ignore[assignment]
+    CommError = RuntimeError  # type: ignore[assignment]
+
+
+def _safe_value(value: Any) -> Any:
+    if isinstance(value, (str, int, float, bool)) or value is None:
+        return value
+    if isinstance(value, (list, tuple)):
+        return [_safe_value(v) for v in value]
+    if isinstance(value, dict):
+        return {str(k): _safe_value(value[k]) for k in sorted(value)}
+    return str(value)
+
+
+def _safe_scope(scope: str | None) -> str:
+    raw = "manual" if scope in (None, "") else str(scope)
+    cleaned = re.sub(r"[^A-Za-z0-9_.-]+", "-", raw).strip("-")
+    return cleaned or "manual"
+
+
+def checkpoint_artifact_name(
+    cfg: Mapping[str, Any], *, backend: str, sweep_id: str | None = None
+) -> str:
+    payload = {k: _safe_value(cfg[k]) for k in sorted(cfg)}
+    scope = _safe_scope(sweep_id)
+    canonical = json.dumps(
+        {"backend": backend, "scope": scope, "cfg": payload},
+        sort_keys=True,
+        separators=(",", ":"),
+    )
+    digest = hashlib.sha1(canonical.encode("utf-8")).hexdigest()[:14]
+    return f"phantom-{backend}-ckpt-{scope}-{digest}"[:128]
+
+
+def _is_missing_artifact_error(exc: Exception) -> bool:
+    if isinstance(exc, CommError):
+        msg = str(exc).lower()
+        return "not found" in msg or "does not exist" in msg
+    return False
+
+
+def download_latest_checkpoint(
+    artifact_name: str, *, file_name: str
+) -> tuple[Path, dict[str, Any]] | None:
+    if not HAS_WANDB or wandb.run is None:
+        return None
+    try:
+        artifact = wandb.run.use_artifact(f"{artifact_name}:latest")
+    except Exception as exc:
+        if _is_missing_artifact_error(exc):
+            return None
+        raise
+    directory = Path(artifact.download())
+    checkpoint_path = directory / file_name
+    if not checkpoint_path.exists():
+        return None
+    metadata = dict(getattr(artifact, "metadata", {}) or {})
+    return checkpoint_path, metadata
+
+
+def _aliases_from_metadata(metadata: dict[str, Any] | None) -> list[str]:
+    aliases = ["latest"]
+    if metadata is None:
+        return aliases
+    if "step" in metadata:
+        try:
+            aliases.append(f"step-{int(metadata['step'])}")
+        except (TypeError, ValueError):
+            pass
+    return aliases
+
+
+def log_checkpoint_bytes(
+    artifact_name: str,
+    *,
+    file_name: str,
+    payload: bytes,
+    metadata: dict[str, Any] | None = None,
+) -> bool:
+    if not HAS_WANDB or wandb.run is None:
+        return False
+    with TemporaryDirectory(prefix="phantom-ckpt-") as tmpdir:
+        path = Path(tmpdir) / file_name
+        path.write_bytes(payload)
+        artifact = wandb.Artifact(
+            name=artifact_name,
+            type="checkpoint",
+            metadata=metadata or {},
+        )
+        artifact.add_file(path.as_posix(), name=file_name)
+        wandb.log_artifact(artifact, aliases=_aliases_from_metadata(metadata))
+    return True
+
+
+def log_checkpoint_file(
+    artifact_name: str,
+    *,
+    file_path: str | Path,
+    artifact_file_name: str,
+    metadata: dict[str, Any] | None = None,
+) -> bool:
+    if not HAS_WANDB or wandb.run is None:
+        return False
+    src = Path(file_path)
+    if not src.exists():
+        return False
+    artifact = wandb.Artifact(
+        name=artifact_name,
+        type="checkpoint",
+        metadata=metadata or {},
+    )
+    artifact.add_file(src.as_posix(), name=artifact_file_name)
+    wandb.log_artifact(artifact, aliases=_aliases_from_metadata(metadata))
+    return True

engine/wrapper.py

@@ -0,0 +1,477 @@
+import gymnasium as gym
+from gymnasium import spaces
+import numpy as np
+from .engine import Limbo, MarketEngine, PricingEngine
+from .lib.render import DashboardRenderer
+from .lib.coi import (
+    compute_uplift_coi,
+    extract_purchases,
+    compute_agent_probability,
+)
+from .lib.behavior import get_transition_models, trajectory_to_events
+from .lib.wrappers import EconomicMetricsWrapper
+from .jax.robust import select_adversarial_alpha_jax, _JAX_OK
+
+
+class _ActionPricingEngine(PricingEngine):
+    def __init__(self, n_products: int, price_bounds: tuple):
+        self._prices = np.full(n_products, price_bounds[0], dtype=float)
+
+    def set_prices(self, prices: np.ndarray):
+        self._prices = np.asarray(prices, dtype=float)
+
+    def act(self, _):
+        return self._prices
+
+
+class PHANTOM(gym.Env):
+    """Gymnasium wrapper for Limbo pricing-market simulation implementing thesis COI framework
+
+    reward = R(p,d) - λ·COI_leak(p,τ') per thesis Section on DR-RL
+    COI_leak uses behavioral divergence to estimate agent probability f(τ')
+    robust inner step: min over alpha in Wasserstein interval around nominal alpha
+    actions are discrete global price-scale moves
+    """
+
+    metadata = {"render_modes": ["human", "ansi"]}
+
+    def __init__(
+        self,
+        n_products: int = 10,
+        alpha: float = 0.3,
+        N: int = 100,
+        human_params: tuple = (50.0, 10.0),
+        agent_params: tuple = (45.0, 15.0),
+        noise_std: float = 1.0,
+        price_bounds: tuple = (10.0, 150.0),
+        lambda_coi: float = 0.1,
+        coi_window: int = 10,
+        robust_radius: float = 0.0,
+        robust_points: int = 5,
+        robust_rollouts: int = 1,
+        info_value: float = 1.0,
+        eta_ux: float = 0.5,
+        reward_profit_weight: float = 1.0,
+        action_levels: int = 9,
+        action_scale_low: float = 0.9,
+        action_scale_high: float = 1.1,
+        max_steps: int = 100,
+        margin_floor: float = 0.05,
+        margin_floor_patience: int = 5,
+        render_mode: str = None,
+    ):
+        super().__init__()
+        self.n_products = n_products
+        self.price_bounds = price_bounds
+        self.lambda_coi = lambda_coi
+        self.coi_window = coi_window
+        self.max_steps = max(1, int(max_steps))
+        self.margin_floor = float(
+            margin_floor
+        )  # terminate if avg margin stays below this for patience steps
+        self.margin_floor_patience = max(1, int(margin_floor_patience))
+        self.render_mode = render_mode
+        self.alpha = float(alpha)
+        self.nominal_alpha = float(alpha)
+        self.N = N
+        self.human_params = human_params
+        self.agent_params = agent_params
+        self.robust_radius = max(0.0, float(robust_radius))
+        self.robust_points = max(1, int(robust_points))
+        self.robust_rollouts = max(1, int(robust_rollouts))
+        self.info_value = float(info_value)
+        self.eta_ux = float(eta_ux)
+        self.reward_profit_weight = float(reward_profit_weight)
+        self.action_levels = max(2, int(action_levels))
+        self._action_scales = np.linspace(
+            float(action_scale_low), float(action_scale_high), self.action_levels
+        )
+
+        self.market = MarketEngine(
+            alpha=alpha,
+            N=N,
+            human_params=human_params,
+            agent_params=agent_params,
+            noise_std=noise_std,
+        )
+        self._platform_stub = _ActionPricingEngine(n_products, price_bounds)
+        self._limbo = Limbo(self._platform_stub, self.market)
+        self._set_market_mix(self.nominal_alpha)
+
+        self.action_space = spaces.Discrete(self.action_levels)
+        self.observation_space = spaces.Dict(
+            {
+                "demand": spaces.Box(
+                    low=0.0, high=100.0, shape=(n_products,), dtype=np.float32
+                ),
+                "prices": spaces.Box(
+                    low=price_bounds[0],
+                    high=price_bounds[1],
+                    shape=(n_products,),
+                    dtype=np.float32,
+                ),
+                "signals": spaces.Box(
+                    low=np.array([0.0, 0.0, 0.0, 0.0], dtype=np.float32),
+                    high=np.array([1.0, 1.0, 1.0, 1.0], dtype=np.float32),
+                    shape=(4,),
+                    dtype=np.float32,
+                ),
+            }
+        )
+
+        self._prices = None
+        self._demand = None
+        self._step_count = 0
+        self._global_step = 0  # monotonic; used as JAX RNG seed across resets
+        self._demand_history = []
+        self._price_history = []
+        self._revenue_history = []
+        self._renderer = None
+        self._initial_episode_prices = None
+        self._trajectories = []  # session trajectories for agent prob calculation
+        self.baseline_prices = np.full(self.n_products, self.price_bounds[0])
+        self.anchor_prices = np.full(
+            self.n_products,
+            float(np.clip(float(self.human_params[0]), *self.price_bounds)),
+        )
+        self.competitive_cap = float(
+            min(self.price_bounds[1], float(np.mean(self.anchor_prices)) * 1.15)
+        )
+        self._low_margin_streak = 0  # consecutive steps below margin_floor
+        self._last_agent_prob = float(self.alpha)
+        self._last_alpha_adv = float(self.alpha)
+
+        # load behavioral models for agent probability estimation
+        try:
+            self._human_trans, self._agent_trans = get_transition_models()
+        except Exception:
+            # fallback if behavioral data unavailable
+            self._human_trans, self._agent_trans = None, None
+
+    def _get_obs(self) -> dict:
+        demand_arr = np.array(
+            [self._demand.get(i, 0.0) for i in range(self.n_products)], dtype=np.float32
+        )
+        signals = np.array(
+            [
+                float(np.clip(self._last_agent_prob, 0.0, 1.0)),
+                float(np.clip(self._last_alpha_adv, 0.0, 1.0)),
+                float(np.clip(self.nominal_alpha, 0.0, 1.0)),
+                float(np.clip(self.robust_radius, 0.0, 1.0)),
+            ],
+            dtype=np.float32,
+        )
+        return {
+            "demand": demand_arr,
+            "prices": self._prices.astype(np.float32),
+            "signals": signals,
+        }
+
+    def _set_market_mix(self, alpha: float):
+        alpha = float(np.clip(alpha, 0.0, 1.0))
+        n_agents = int(self.N * alpha)
+        self.alpha = alpha
+        self.market.alpha = alpha
+        self.market.Nagents = n_agents
+        self.market.Nhumans = self.N - n_agents
+
+    def _decode_action(self, action) -> np.ndarray:
+        prev = self._prices
+        base = self.anchor_prices
+
+        def _blend(target: np.ndarray) -> np.ndarray:
+            if prev is None:
+                lower = float(self.price_bounds[0])
+                return np.clip(target, lower, self.competitive_cap)
+            blended = 0.75 * np.asarray(prev, dtype=float) + 0.25 * target
+            lower = float(self.price_bounds[0])
+            return np.clip(blended, lower, self.competitive_cap)
+
+        if np.isscalar(action):
+            idx = int(np.clip(int(action), 0, self.action_levels - 1))
+            target = base * self._action_scales[idx]
+            return _blend(target)
+        a = np.asarray(action)
+        if a.size == 1:
+            idx = int(np.clip(int(a.reshape(-1)[0]), 0, self.action_levels - 1))
+            target = base * self._action_scales[idx]
+            return _blend(target)
+        lower = float(self.price_bounds[0])
+        return np.clip(a.astype(float), lower, self.competitive_cap)
+
+    def _compute_agent_prob(self, trajectories=None) -> float:
+        trajectories = (
+            self.market.last_trajectories if trajectories is None else trajectories
+        )
+        if not trajectories or self._human_trans is None or self._agent_trans is None:
+            return float(self.market.alpha)
+        probs = []
+        for traj in trajectories:
+            events = trajectory_to_events(traj)
+            if len(events) < 2:
+                continue
+            probs.append(
+                compute_agent_probability(events, self._human_trans, self._agent_trans)
+            )
+        return float(np.mean(probs)) if probs else float(self.market.alpha)
+
+    def _compute_reward(
+        self, prices: np.ndarray, demand: dict, agent_prob: float, trajectories: list
+    ) -> tuple[float, dict]:
+        demand_arr = np.array(
+            [demand.get(i, 0.0) for i in range(self.n_products)], dtype=float
+        )
+        revenue = float(np.dot(prices, demand_arr))
+        floor_cost = float(np.dot(self.baseline_prices, demand_arr))
+        profit = revenue - floor_cost
+        purchases = extract_purchases(trajectories)
+        coi_mix = compute_uplift_coi(prices, purchases, self.baseline_prices)
+
+        coi_leakage = float(agent_prob * self.info_value)
+        info_budget = max(floor_cost, 1.0)
+        coi_penalty = self.lambda_coi * coi_leakage * info_budget
+
+        if len(self._price_history) > 0:
+            prev_prices = np.asarray(self._price_history[-1], dtype=float)
+            rel_change = (prices - prev_prices) / np.maximum(prev_prices, 1.0)
+            volatility = float(np.mean(np.abs(rel_change)))
+            upward_volatility = float(np.mean(np.clip(rel_change, 0.0, None)))
+        else:
+            volatility = 0.0
+            upward_volatility = 0.0
+        ux_penalty = self.eta_ux * info_budget * (volatility + 0.5 * upward_volatility)
+
+        competitive_anchor = float(np.mean(self.anchor_prices))
+        price_ratio = prices / max(competitive_anchor, 1.0)
+        supra_excess = np.clip(price_ratio - 1.15, 0.0, None)
+        supra_penalty = 4.0 * info_budget * float(np.mean(np.square(supra_excess)))
+        supra_share = float(np.mean(supra_excess > 0.0))
+
+        reward_revenue = self.reward_profit_weight * profit
+        reward = reward_revenue - coi_penalty - ux_penalty - supra_penalty
+
+        return reward, {
+            "revenue": revenue,
+            "cost_floor": floor_cost,
+            "profit": profit,
+            "coi_mix": float(coi_mix),
+            "coi_base": 0.0,
+            "coi_leakage": coi_leakage,
+            "coi_penalty": coi_penalty,
+            "coi_info_budget": info_budget,
+            "ux_penalty": ux_penalty,
+            "volatility": volatility,
+            "upward_volatility": upward_volatility,
+            "supra_penalty": supra_penalty,
+            "supra_share": supra_share,
+            "competitive_anchor": competitive_anchor,
+            "reward_revenue": reward_revenue,
+            "reward_total": reward,
+        }
+
+    def _alpha_candidates(self) -> np.ndarray:
+        if self.robust_radius <= 0.0 or self.robust_points == 1:
+            return np.array([self.nominal_alpha], dtype=float)
+        lo = max(0.0, self.nominal_alpha - self.robust_radius)
+        hi = min(1.0, self.nominal_alpha + self.robust_radius)
+        return np.linspace(lo, hi, self.robust_points)
+
+    def _evaluate_candidate(self, alpha: float, prices: np.ndarray) -> float:
+        self._set_market_mix(alpha)
+        rewards = []
+        for _ in range(self.robust_rollouts):
+            demand = self.market.act(prices)
+            trajectories = list(self.market.last_trajectories)
+            agent_prob = self._compute_agent_prob(trajectories)
+            reward, _ = self._compute_reward(prices, demand, agent_prob, trajectories)
+            rewards.append(float(reward))
+        return float(np.mean(rewards)) if rewards else 0.0
+
+    def _select_adversarial_alpha(self, prices: np.ndarray) -> float:
+        """inner robust step: pick worst-case alpha from the ambiguity interval.
+
+        when JAX is available and robust_rollouts==1 we use a vmapped pass over
+        all K candidates in a single call (no Python loop, no market.act overhead).
+        the JAX path approximates demand as the mixed closed-form d(p;theta) signal
+        rather than running full trajectory sampling, which is accurate for the
+        alpha-selection decision while being dramatically cheaper.
+
+        when robust_rollouts>1 or JAX is unavailable we fall back to the sequential
+        market.act() loop so behavior is identical to the original implementation.
+        """
+        candidates = self._alpha_candidates()
+        if len(candidates) == 1:
+            return float(candidates[0])
+
+        if _JAX_OK and self.robust_rollouts == 1:
+            best_alpha, _ = select_adversarial_alpha_jax(
+                candidates=candidates,
+                prices=prices,
+                human_params=self.market.human_params,
+                agent_params=self.market.agent_params,
+                noise_std=self.market.noise_std,
+                baseline_prices=self.baseline_prices,
+                lambda_coi=self.lambda_coi,
+                info_value=self.info_value,
+                reward_profit_weight=self.reward_profit_weight,
+                rng_seed=self._global_step,
+            )
+            return best_alpha
+
+        # fallback: full trajectory-based sequential evaluation
+        evaluations = [
+            (float(alpha), self._evaluate_candidate(float(alpha), prices))
+            for alpha in candidates
+        ]
+        best_alpha, _ = min(evaluations, key=lambda x: x[1])
+        return best_alpha
+
+    def _record_history(self):
+        demand_arr = np.array(
+            [self._demand.get(i, 0.0) for i in range(self.n_products)]
+        )
+        self._demand_history.append(demand_arr)
+        self._price_history.append(self._prices.copy())
+        self._revenue_history.append(np.sum(self._prices * demand_arr))
+
+    def reset(self, seed=None, options=None):
+        super().reset(seed=seed)
+        self._set_market_mix(self.nominal_alpha)
+        self._limbo.reset()
+        self._prices = np.random.uniform(*self.price_bounds, size=self.n_products)
+        self._platform_stub.set_prices(self._prices)
+        self._limbo.step()
+        self._demand = self._limbo.step()
+        self._initial_episode_prices = self._prices.copy()
+        self._step_count = 0
+        self._low_margin_streak = 0
+        self._demand_history, self._price_history, self._revenue_history = [], [], []
+        self._trajectories = list(getattr(self.market, "last_trajectories", []))
+        self._last_agent_prob = float(self.nominal_alpha)
+        self._last_alpha_adv = float(self.nominal_alpha)
+        self._record_history()
+        return self._get_obs(), {}
+
+    def step(self, action):
+        self._prices = self._decode_action(action)
+        alpha_adv = self._select_adversarial_alpha(self._prices)
+        self._global_step += 1  # always increment; JAX path may have already done so
+        self._set_market_mix(alpha_adv)
+        self._platform_stub.set_prices(self._prices)
+        self._step_count += 1
+
+        self._demand = self.market.act(self._prices)
+        trajectories = list(self.market.last_trajectories)
+        agent_prob = self._compute_agent_prob(trajectories)
+        self._trajectories.extend(trajectories)
+        self._last_agent_prob = float(agent_prob)
+        self._last_alpha_adv = float(alpha_adv)
+
+        reward, metrics = self._compute_reward(
+            self._prices, self._demand, agent_prob, trajectories
+        )
+        self._record_history()
+
+        # soft early termination when margin collapses for too long
+        avg_margin = float(np.mean(self._prices) - self.price_bounds[0]) / max(
+            float(np.mean(self._prices)), 1e-6
+        )
+        if avg_margin < self.margin_floor:
+            self._low_margin_streak += 1
+        else:
+            self._low_margin_streak = 0
+        margin_collapsed = self._low_margin_streak >= self.margin_floor_patience
+        terminated = self._step_count >= self.max_steps or margin_collapsed
+
+        info = {
+            "step": self._step_count,
+            "agent_prob": agent_prob,
+            "alpha_adv": float(alpha_adv),
+            "alpha_nominal": float(self.nominal_alpha),
+            "wasserstein_radius": float(self.robust_radius),
+            "robust_rollouts": int(self.robust_rollouts),
+            **metrics,
+            "raw_revenue": np.sum(
+                self._prices
+                * np.array([self._demand.get(i, 0.0) for i in range(self.n_products)])
+            ),
+        }
+        return self._get_obs(), reward, terminated, False, info
+
+    def _compute_elasticity(self) -> np.ndarray:
+        """point elasticity: e = (dQ/dP) * (P/Q) via finite differences, clipped to [-5, 5]"""
+        if len(self._price_history) < 2:
+            return np.zeros(self.n_products)
+        p, q = np.array(self._price_history), np.array(self._demand_history)
+        dp, dq = np.diff(p, axis=0), np.diff(q, axis=0)
+        valid = np.abs(dp) > 0.5
+        with np.errstate(divide="ignore", invalid="ignore"):
+            elasticity = np.where(
+                valid, (dq / dp) * (p[:-1] / np.maximum(q[:-1], 1.0)), 0.0
+            )
+            elasticity = np.nan_to_num(np.clip(elasticity, -5.0, 5.0), nan=0.0)
+        return (
+            np.mean(elasticity, axis=0)
+            if len(elasticity) > 0
+            else np.zeros(self.n_products)
+        )
+
+    def render(self):
+        if self.render_mode == "human":
+            if self._renderer is None:
+                self._renderer = DashboardRenderer()
+            self._renderer.render(self)
+        elif self.render_mode == "ansi":
+            return (
+                f"step={self._step_count}, prices={self._prices}, demand={self._demand}"
+            )
+        return None
+
+    def close(self):
+        if self._renderer:
+            self._renderer.close()
+            self._renderer = None
+
+
+if __name__ == "__main__":
+    import wandb
+    from .lib import MetricsCallback
+
+    class RandomPolicy:
+        """Minimal SB3-compatible random policy for baseline testing."""
+
+        def __init__(self, env):
+            self.env = env
+            self.num_timesteps = 0
+
+        def learn(self, total_timesteps, callback=None):
+            callback.model = self
+            callback.num_timesteps = 0
+            callback.locals = {}
+            callback.on_training_start({}, {})
+
+            obs, _ = self.env.reset()
+            for step in range(total_timesteps):
+                action = self.env.action_space.sample()
+                obs, reward, term, trunc, info = self.env.step(action)
+                self.num_timesteps = step + 1
+                callback.num_timesteps = self.num_timesteps
+                callback.locals = {"infos": [info]}
+                callback.on_step()
+                if term or trunc:
+                    callback.on_rollout_end()
+                    obs, _ = self.env.reset()
+            return self
+
+        def predict(self, obs, **kwargs):
+            return self.env.action_space.sample(), None
+
+    wandb.init(project="capstone", config={"policy": "random", "alpha": 0.3})
+    env = EconomicMetricsWrapper(PHANTOM(n_products=15, alpha=0.3, render_mode=None))
+
+    model = RandomPolicy(env)
+    model.learn(total_timesteps=1000, callback=MetricsCallback())
+
+    print(f"Episode revenue: {env.episode_revenue:.1f}")
+    wandb.finish()
+    env.close()

experiments/README.md

@@ -0,0 +1,8 @@
+
+# Products
+# Agents
+# Pipeline
+
+Our pipeline technically should follow principles in a style like this:
+- Each step should be defined as an inheriting child of an scikit pipeline step, the granularity of the steps is dictated by the following: a step should be a transformation, augmentation or computation independently, no single stage should run multiple in-itself. This way we can modularize properly all the components and track properly in airflow. A stage can be defined as an sklearn step but then must be transalted to a function that takes the context in our DAG of airflow. All parametrization must be done via contexts.
+

experiments/agents/agent.py

@@ -38,7 +38,10 @@ def get_agent(agent_type: AgentTypes,  **kwargs) -> Agent:
 
 if __name__ == "__main__":
     import asyncio
-    JTBD= "Name all the products on this site and try to find out more about each product by clicking into them (they might not open)"
-    agent = get_agent(AgentTypes.GENERIC_BROWSER_USE_AGENT, goal=JTBD, url="http://localhost:3000/products", timeout=300)
+    JTBD= "Find me the cheapest room in Madrid for 2 people in the next two days, review each hotel room in detail and then add it to cart."
+    agent = get_agent(AgentTypes.GENERIC_BROWSER_USE_AGENT,
+                      goal=JTBD,
+                      url="http://localhost:3000/start-task?uuid=d10f5ab3-a7b7-4e97-8d94-ab06f1537c0a",
+                      timeout=300)
     R=asyncio.run(agent.act())
     print(R)

experiments/agents/run.py

@@ -0,0 +1,117 @@
+from supabase import create_client, Client
+import os
+import random
+import asyncio
+import json
+from dotenv import load_dotenv
+
+from experiments.agents.agent import get_agent, AgentTypes
+from lib.kafka_client import get_interactions
+
+load_dotenv()
+
+RESULTS="/home/velocitatem/Documents/Projects/PHANTOM/experiments/agents/collected_data/"
+
+client = create_client(
+    os.getenv("NEXT_PUBLIC_SUPABASE_URL"),
+    os.getenv("NEXT_PUBLIC_SUPABASE_ANON_KEY")
+)
+def pick_random_task():
+    mode = 'hotel'
+    tasks = client.table("tasks").select("*").execute().data
+    if mode == 'hotel':
+        # drop all that have 'flight' in the description
+        tasks = [task for task in tasks if 'flight' not in task['task_description'].lower()]
+    return random.choice(tasks) if tasks else None
+
+def clear_kafka_data():
+    """Delete and recreate Kafka topics to clear all data"""
+    from kafka.admin import KafkaAdminClient, NewTopic
+    from kafka.errors import UnknownTopicOrPartitionError
+    import time
+
+    kafka_host = os.getenv('KAFKA_HOST', 'localhost')
+    kafka_port = os.getenv('KAFKA_PORT', '9092')
+    broker = f'{kafka_host}:{kafka_port}'
+
+    admin = KafkaAdminClient(bootstrap_servers=broker)
+    topics = ['user-interactions', 'price-logs']
+
+    try:
+        admin.delete_topics(topics, timeout_ms=5000)
+        print(f"Deleted topics: {topics}")
+        time.sleep(2)
+    except UnknownTopicOrPartitionError:
+        print("Topics don't exist, skipping delete")
+    except Exception as e:
+        print(f"Error deleting topics: {e}")
+
+    new_topics = [
+        NewTopic(name='user-interactions', num_partitions=3, replication_factor=1),
+        NewTopic(name='price-logs', num_partitions=3, replication_factor=1)
+    ]
+
+    try:
+        admin.create_topics(new_topics=new_topics, validate_only=False)
+        print(f"Recreated topics: {topics}")
+    except Exception as e:
+        print(f"Error creating topics: {e}")
+    finally:
+        admin.close()
+
+def create_new_experiment(task_id):
+    import uuid
+    subject_name = f"agent_{str(uuid.uuid4())[:8]}"
+    experiment = {
+        "subject_name": subject_name,
+        "xp_human_only": False,
+        "xp_market_mode": "hotel",
+        "xp_task_id": task_id,
+    }
+    response = client.table("experiments").insert(experiment).execute()
+    return response.data[0] if response.data else None
+
+if __name__ == "__main__":
+    clear_kafka_data()
+
+    task = pick_random_task()
+    if not task:
+        print("No tasks available")
+        exit(1)
+
+    experiment = create_new_experiment(task['id'])
+    exp_id = experiment['id']
+    exp_dir = f"{RESULTS}{exp_id}"
+    os.makedirs(exp_dir, exist_ok=True)
+
+    # construct experiment URL with uuid param
+    base_url = os.getenv('NEXT_PUBLIC_API_BASE', 'http://localhost:3000')
+    agent_url = f"{base_url}/start-task?uuid={exp_id}"
+
+    print(f"Created experiment {exp_id} for task {task['id']}")
+    print(f"Agent will interact with: {agent_url}")
+
+    # instantiate and run agent
+    agent = get_agent(
+        AgentTypes.GENERIC_BROWSER_USE_AGENT,
+        goal=task['task_description'],
+        url=agent_url,
+        timeout=300,
+        headless=True
+    )
+
+    result = asyncio.run(agent.act())
+    print(f"Agent result: {result}")
+
+    # export interaction and price data from kafka
+    interactions = get_interactions(topic='user-interactions', timeout_ms=3000)
+    prices = get_interactions(topic='price-logs', timeout_ms=3000)
+
+    with open(f"{exp_dir}/int.json", 'w') as f:
+        json.dump(interactions, f, indent=2)
+
+    with open(f"{exp_dir}/price.json", 'w') as f:
+        json.dump(prices, f, indent=2)
+
+    print(f"Experiment {exp_id} completed.")
+    print(f"Exported {len(interactions)} interactions and {len(prices)} price logs to {exp_dir}")

experiments/airflow/dags/ml_training_pipeline.py

@@ -0,0 +1,115 @@
+from airflow import DAG, Dataset
+from airflow.decorators import task
+from airflow.utils.dates import days_ago
+from datetime import timedelta
+import pandas as pd
+import logging
+import sys
+import pickle
+
+sys.path.insert(0, '/opt/airflow')
+
+from procesing.context import PipelineContext
+from procesing.providers import SupabaseProvider, BackendAPIProvider
+from procesing.steps import (
+    FetchInteractionsStep,
+    ValidateDataStep,
+    ExtractSessionFeaturesStep,
+    JoinLabelsStep,
+)
+
+TRAINING_DATASET = Dataset('phantom://ml/training-data')
+
+DEFAULT_ARGS = {
+    'owner': 'phantom-research',
+    'depends_on_past': False,
+    'email_on_failure': False,
+    'email_on_retry': False,
+    'retries': 2,
+    'retry_delay': timedelta(minutes=5),
+}
+
+
+class CompositeProvider(SupabaseProvider, BackendAPIProvider):
+    def __init__(self):
+        SupabaseProvider.__init__(self)
+        BackendAPIProvider.__init__(self)
+
+
+def _get_context(store_mode: str = 'hotel') -> PipelineContext:
+    return PipelineContext(provider=CompositeProvider(), store_mode=store_mode)
+
+
+with DAG(
+    'ml_training_pipeline',
+    default_args=DEFAULT_ARGS,
+    description='ML training data pipeline: fetch -> validate -> extract features -> label -> publish',
+    schedule=None,
+    start_date=days_ago(1),
+    catchup=False,
+    max_active_runs=1,
+    tags=['ml', 'training', 'features', 'research'],
+) as dag:
+
+    @task
+    def fetch_interactions(**kwargs) -> bytes:
+        dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+        ctx = _get_context(dag_conf.get('store_mode', 'hotel'))
+        df = FetchInteractionsStep(ctx).transform(None)
+        logging.info(f"Fetched {len(df)} interactions, {df['sessionId'].nunique()} sessions")
+        return pickle.dumps(df)
+
+    @task
+    def validate_data(raw_data: bytes, **kwargs) -> bytes:
+        df = pickle.loads(raw_data)
+        dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+        ctx = _get_context(dag_conf.get('store_mode', 'hotel'))
+        validated = ValidateDataStep(ctx).transform(df)
+        report = ctx.get_cached('validation_report') or {}
+        logging.info(f"Validation: {report.get('status')}, {report.get('sessions', 0)} sessions")
+        return pickle.dumps(validated)
+
+    @task
+    def extract_session_features(validated_data: bytes, **kwargs) -> bytes:
+        df = pickle.loads(validated_data)
+        if df.empty:
+            logging.warning("Empty input, skipping feature extraction")
+            return pickle.dumps(pd.DataFrame())
+        dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+        ctx = _get_context(dag_conf.get('store_mode', 'hotel'))
+        features = ExtractSessionFeaturesStep(ctx).transform(df)
+        logging.info(f"Extracted {len(features.columns)} features for {len(features)} sessions")
+        return pickle.dumps(features)
+
+    @task
+    def join_labels(features_data: bytes, **kwargs) -> bytes:
+        features_df = pickle.loads(features_data)
+        if features_df.empty:
+            logging.warning("Empty features, skipping label join")
+            return pickle.dumps(pd.DataFrame())
+        dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+        ctx = _get_context(dag_conf.get('store_mode', 'hotel'))
+        labeled = JoinLabelsStep(ctx).transform(features_df)
+        n_agents = labeled['is_agent'].sum() if 'is_agent' in labeled.columns else 0
+        logging.info(f"Labeled {len(labeled)} sessions: {n_agents} agents")
+        return pickle.dumps(labeled)
+
+    @task(outlets=[TRAINING_DATASET])
+    def publish_training_data(labeled_data: bytes, **kwargs) -> dict:
+        labeled_df = pickle.loads(labeled_data)
+        if labeled_df.empty:
+            return {'status': 'skipped', 'reason': 'empty_data'}
+        dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+        return {
+            'status': 'success',
+            'n_sessions': len(labeled_df),
+            'n_features': len([c for c in labeled_df.columns if c not in ['sessionId', 'experimentId', 'is_agent']]),
+            'store_mode': dag_conf.get('store_mode', 'hotel'),
+            'timestamp': pd.Timestamp.now().isoformat(),
+        }
+
+    raw = fetch_interactions()
+    validated = validate_data(raw)
+    features = extract_session_features(validated)
+    labeled = join_labels(features)
+    publish_training_data(labeled)

experiments/airflow/dags/session_pricing_pipeline.py

@@ -0,0 +1,269 @@
+"""
+Session-Aware Pricing DAG
+THIS implements the core pricing computation (policy layer).
+
+Flow: τ → θ̂ → D → p*
+  1. Fetch recent sessions from Kafka (last 10 active)
+  2. Extract features per session (τ → θ̂)
+  3. Map features to demand proxy (θ̂ → D)
+  4. Compute optimal prices (D → p*)
+  5. Write to Redis session:{sessionId}:prices
+
+Scheduled: every 1 minute when enabled
+"""
+from airflow import DAG
+from airflow.operators.python import PythonOperator
+from airflow.utils.dates import days_ago
+from datetime import timedelta
+import pandas as pd
+import numpy as np
+import logging
+import sys
+import pickle
+
+sys.path.insert(0, '/opt/airflow')
+
+from procesing.context import PipelineContext
+from procesing.providers import SupabaseProvider, BackendAPIProvider
+from procesing.steps.session import ExtractSessionFeaturesStep
+from procesing.pricers.simple import SimpleSurgePricer, session_features_to_demand
+from procesing.pricing import StateSpace
+from lib.model_registry import ModelRegistry
+
+DEFAULT_ARGS = {
+    'owner': 'phantom-research',
+    'depends_on_past': False,
+    'email_on_failure': False,
+    'email_on_retry': False,
+    'retries': 1,
+    'retry_delay': timedelta(seconds=30),
+}
+
+
+class CompositeProvider(SupabaseProvider, BackendAPIProvider):
+    def __init__(self):
+        SupabaseProvider.__init__(self)
+        BackendAPIProvider.__init__(self)
+
+
+def _get_context(store_mode: str = 'hotel') -> PipelineContext:
+    return PipelineContext(provider=CompositeProvider(), store_mode=store_mode)
+
+
+def fetch_recent_sessions(**kwargs):
+    """
+    Task: Fetch last N active sessions from Kafka.
+    Returns: DataFrame of interaction events for recent sessions.
+    """
+    dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+    store_mode = dag_conf.get('store_mode', 'hotel')
+    session_limit = dag_conf.get('session_limit', 10)
+
+    ctx = _get_context(store_mode)
+    provider = ctx.provider
+
+    # fetch all recent interactions from Kafka
+    try:
+        interactions_df = provider.fetch_kafka_topic("user-interactions")
+    except Exception as e:
+        logging.error(f"Failed to fetch interactions: {e}")
+        kwargs['ti'].xcom_push(key='sessions_data', value=pickle.dumps(pd.DataFrame()))
+        return 0
+
+    if interactions_df.empty or 'sessionId' not in interactions_df.columns:
+        kwargs['ti'].xcom_push(key='sessions_data', value=pickle.dumps(pd.DataFrame()))
+        return 0
+
+    # identify last N active sessions (most recent by event count)
+    recent_sessions = interactions_df['sessionId'].value_counts().head(session_limit).index.tolist()
+
+    # filter to only those sessions
+    filtered_df = interactions_df[interactions_df['sessionId'].isin(recent_sessions)].copy()
+
+    kwargs['ti'].xcom_push(key='sessions_data', value=pickle.dumps(filtered_df))
+    kwargs['ti'].xcom_push(key='session_ids', value=recent_sessions)
+
+    logging.info(f"Fetched {len(filtered_df)} events for {len(recent_sessions)} sessions")
+    return len(recent_sessions)
+
+
+def extract_session_features(**kwargs):
+    """
+    Task: Extract behavioral features from session trajectories.
+    THIS implements τ → θ̂ transformation.
+    """
+    ti = kwargs['ti']
+    sessions_df = pickle.loads(ti.xcom_pull(key='sessions_data'))
+
+    if sessions_df.empty:
+        ti.xcom_push(key='session_features', value=pickle.dumps(pd.DataFrame()))
+        return 0
+
+    dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+    ctx = _get_context(dag_conf.get('store_mode', 'hotel'))
+
+    # extract features using vectorized pipeline
+    feature_extractor = ExtractSessionFeaturesStep(ctx)
+    features_df = feature_extractor.transform(sessions_df)
+
+    ti.xcom_push(key='session_features', value=pickle.dumps(features_df))
+
+    logging.info(f"Extracted {len(features_df.columns)} features for {len(features_df)} sessions")
+    logging.info(f"Feature columns: {list(features_df.columns)}")
+    logging.info(f"Sample features (first session):\n{features_df.iloc[0].to_dict()}")
+
+    return len(features_df)
+
+
+def compute_session_prices(**kwargs):
+    """
+    Task: Compute optimal prices for each session.
+    THIS implements θ̂ → D → p* transformation.
+    """
+    ti = kwargs['ti']
+    features_df = pickle.loads(ti.xcom_pull(key='session_features'))
+
+    if features_df.empty:
+        ti.xcom_push(key='price_results', value=pickle.dumps({}))
+        return 0
+
+    dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+    store_mode = dag_conf.get('store_mode', 'hotel')
+    ctx = _get_context(store_mode)
+
+    # fetch product catalog for base prices
+    products_df = ctx.provider.fetch_products(store_mode)
+    if products_df.empty:
+        logging.error("No products found in catalog")
+        ti.xcom_push(key='price_results', value=pickle.dumps({}))
+        return 0
+
+    products_df['base_price'] = products_df['metadata'].apply(
+        lambda m: m.get('base_price', 100.0) if isinstance(m, dict) else 100.0
+    )
+
+    # initialize pricing model
+    pricer = SimpleSurgePricer(
+        high_threshold=dag_conf.get('high_threshold', 10),
+        low_threshold=dag_conf.get('low_threshold', 2),
+        surge_multiplier=dag_conf.get('surge_multiplier', 1.15),
+        discount_multiplier=dag_conf.get('discount_multiplier', 0.95)
+    )
+    pricer.fit(products_df)
+
+    # compute prices per session
+    price_results = {}
+    n_products = len(products_df)
+
+    logging.info(f"Starting price computation for {len(features_df)} sessions, {n_products} products")
+    logging.info(f"Pricer config: high_thresh={pricer.high_threshold}, low_thresh={pricer.low_threshold}, surge_mult={pricer.surge_multiplier}")
+
+    for idx, session_row in features_df.iterrows():
+        session_id = session_row.get('sessionId')
+        if not session_id:
+            continue
+
+        # map features to demand proxy (θ̂ → D)
+        session_features_single = pd.DataFrame([session_row])
+        demand_proxy = session_features_to_demand(session_features_single)
+
+        logging.info(f"[Session {session_id}] Features → Demand: {demand_proxy:.2f}")
+        logging.info(f"[Session {session_id}] Key features: velocity={session_row.get('interaction_velocity', 0):.2f}, cart_ratio={session_row.get('cart_to_view_ratio', 0):.2f}, item_views={session_row.get('item_views', 0)}")
+
+        # build state space
+        state_space = StateSpace(
+            demand=np.full(n_products, demand_proxy),  # broadcast session demand to all products
+            prices=products_df['base_price'].values,
+            session_features=session_features_single
+        )
+
+        # compute optimal prices (D → p*)
+        optimal_prices = pricer.predict(state_space)
+
+        base_avg = products_df['base_price'].mean()
+        optimal_avg = optimal_prices.mean()
+        price_change_pct = ((optimal_avg - base_avg) / base_avg) * 100
+
+        logging.info(f"[Session {session_id}] Price adjustment: base_avg={base_avg:.2f}, optimal_avg={optimal_avg:.2f}, change={price_change_pct:+.1f}%")
+
+        # store as dict {productId: price}
+        price_map = {
+            str(products_df.iloc[i]['id']): float(optimal_prices[i])
+            for i in range(n_products)
+        }
+
+        price_results[session_id] = price_map
+
+    ti.xcom_push(key='price_results', value=pickle.dumps(price_results))
+
+    logging.info(f"Computed prices for {len(price_results)} sessions, {n_products} products each")
+    return len(price_results)
+
+
+def publish_to_registry(**kwargs):
+    """
+    Task: Write session prices to Redis registry.
+    THIS is the write path: prices → session:{sessionId}:prices
+    """
+    ti = kwargs['ti']
+    price_results = pickle.loads(ti.xcom_pull(key='price_results'))
+
+    if not price_results:
+        logging.warning("No prices to publish")
+        return 0
+
+    registry = ModelRegistry()
+    ttl = kwargs.get('dag_run').conf.get('ttl', 1800) if kwargs.get('dag_run') and kwargs.get('dag_run').conf else 1800
+
+    published_count = 0
+    for session_id, price_map in price_results.items():
+        registry.set_session_prices(session_id, price_map, ttl=ttl)
+        published_count += 1
+
+    logging.info(f"Published prices for {published_count} sessions to registry (TTL={ttl}s)")
+
+    return {
+        'sessions_published': published_count,
+        'products_per_session': len(next(iter(price_results.values()))) if price_results else 0,
+        'status': 'success'
+    }
+
+
+# DAG definition
+with DAG(
+    'session_pricing_pipeline',
+    default_args=DEFAULT_ARGS,
+    description='Session-aware pricing: extract features → compute prices → publish to registry',
+    schedule_interval='*/1 * * * *',  # every 1 minute
+    start_date=days_ago(1),
+    catchup=False,
+    max_active_runs=1,
+    tags=['pricing', 'session-aware', 'research', 'real-time'],
+) as dag:
+
+    t_fetch_sessions = PythonOperator(
+        task_id='fetch_recent_sessions',
+        python_callable=fetch_recent_sessions,
+        provide_context=True,
+    )
+
+    t_extract_features = PythonOperator(
+        task_id='extract_session_features',
+        python_callable=extract_session_features,
+        provide_context=True,
+    )
+
+    t_compute_prices = PythonOperator(
+        task_id='compute_session_prices',
+        python_callable=compute_session_prices,
+        provide_context=True,
+    )
+
+    t_publish = PythonOperator(
+        task_id='publish_to_registry',
+        python_callable=publish_to_registry,
+        provide_context=True,
+    )
+
+    # linear dependency: fetch → extract → compute → publish
+    t_fetch_sessions >> t_extract_features >> t_compute_prices >> t_publish

experiments/airflow/dags/surge_pricing_factory.py

@@ -0,0 +1,220 @@
+from pandas.core.algorithms import factorize_array
+from airflow import DAG
+from airflow.operators.python import PythonOperator
+from airflow.utils.dates import days_ago
+from datetime import timedelta
+import pandas as pd
+import logging
+import sys
+import pickle
+
+sys.path.insert(0, '/opt/airflow')
+
+from procesing.context import PipelineContext
+from procesing.providers import SupabaseProvider, BackendAPIProvider
+from procesing.steps import (
+    FetchInteractionsStep,
+    FetchPriceLogsStep,
+    ComputeDemandStep,
+    AggregatePriceLogsStep,
+    JoinProductFeaturesStep,
+)
+from procesing.pricers.simple import SimpleSurgePricer
+
+DEFAULT_ARGS = {
+    'owner': 'phantom-research',
+    'depends_on_past': False,
+    'email_on_failure': False,
+    'email_on_retry': False,
+    'retries': 2,
+    'retry_delay': timedelta(minutes=5),
+}
+
+class CompositeProvider(SupabaseProvider, BackendAPIProvider):
+    def __init__(self):
+        SupabaseProvider.__init__(self)
+        BackendAPIProvider.__init__(self)
+
+def _get_provider():
+    return CompositeProvider()
+
+def _make_task_callables(store_mode: str):
+    """Generate task callables bound to a specific store_mode."""
+
+    def get_context(**kwargs):
+        return PipelineContext(provider=_get_provider(), store_mode=store_mode)
+
+    def fetch_interactions(**kwargs):
+        ctx = get_context(**kwargs)
+        df = FetchInteractionsStep(ctx).transform(None)
+        kwargs['ti'].xcom_push(key='interactions_raw', value=pickle.dumps(df))
+        logging.info(f"[{store_mode}] Fetched {len(df)} interaction records")
+        return len(df)
+
+    def fetch_price_logs(**kwargs):
+        ctx = get_context(**kwargs)
+        df = FetchPriceLogsStep(ctx).transform(None)
+        kwargs['ti'].xcom_push(key='price_logs_raw', value=pickle.dumps(df))
+        logging.info(f"[{store_mode}] Fetched {len(df)} price records")
+        return len(df)
+
+    def compute_demand(**kwargs):
+        ti = kwargs['ti']
+        df = pickle.loads(ti.xcom_pull(key='interactions_raw'))
+        ctx = get_context(**kwargs)
+        demand_df = ComputeDemandStep(ctx).transform(df)
+        ti.xcom_push(key='demand_data', value=pickle.dumps(demand_df))
+        logging.info(f"[{store_mode}] Computed demand for {len(demand_df)} products")
+        return len(demand_df)
+
+    def aggregate_price_logs(**kwargs):
+        ti = kwargs['ti']
+        df = pickle.loads(ti.xcom_pull(key='price_logs_raw'))
+        ctx = get_context(**kwargs)
+        price_df = AggregatePriceLogsStep(ctx).transform(df)
+        ti.xcom_push(key='price_data', value=pickle.dumps(price_df))
+        logging.info(f"[{store_mode}] Aggregated price logs for {len(price_df)} products")
+        return len(price_df)
+
+    def join_product_features(**kwargs):
+        ti = kwargs['ti']
+        demand_df = pickle.loads(ti.xcom_pull(key='demand_data'))
+        price_df = pickle.loads(ti.xcom_pull(key='price_data'))
+        ctx = get_context(**kwargs)
+        joined_df = JoinProductFeaturesStep(ctx).transform((demand_df, price_df))
+        ti.xcom_push(key='product_features', value=pickle.dumps(joined_df))
+        logging.info(f"[{store_mode}] Joined features for {len(joined_df)} products")
+        return len(joined_df)
+
+    def apply_surge_pricing(**kwargs):
+        ti = kwargs['ti']
+        product_features = pickle.loads(ti.xcom_pull(key='product_features'))
+        dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+
+        data = product_features.rename(columns={'demand_score': 'demand'})
+        surge_pricer = SimpleSurgePricer(
+            high_threshold=dag_conf.get('high_threshold', 10),
+            low_threshold=dag_conf.get('low_threshold', 2),
+            surge_multiplier=dag_conf.get('surge_multiplier', 1.2),
+            discount_multiplier=dag_conf.get('discount_multiplier', 0.9)
+        )
+        surge_pricer.fit(data)
+        data['optimal_price'] = surge_pricer.predict()
+
+        prices_df = data[['productId', 'price', 'base_price', 'optimal_price', 'demand']].rename(columns={
+            'price': 'current_price', 'demand': 'demand_score'
+        })
+        ti.xcom_push(key='predicted_prices', value=pickle.dumps(prices_df))
+        logging.info(f"[{store_mode}] Applied surge pricing for {len(prices_df)} products")
+        return len(prices_df)
+
+    def publish_results(**kwargs):
+        ti = kwargs['ti']
+        prices_df = pickle.loads(ti.xcom_pull(key='predicted_prices'))
+        from lib.model_registry import ModelRegistry
+
+        registry = ModelRegistry()
+        dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+
+        metadata = {
+            'timestamp': pd.Timestamp.now().isoformat(),
+            'store_mode': store_mode,
+            'dag_run_id': kwargs['dag_run'].run_id if kwargs.get('dag_run') else 'manual',
+            'pricing_method': 'surge',
+            'high_threshold': dag_conf.get('high_threshold', 10),
+            'low_threshold': dag_conf.get('low_threshold', 2),
+            'surge_multiplier': dag_conf.get('surge_multiplier', 1.2),
+            'discount_multiplier': dag_conf.get('discount_multiplier', 0.9)
+        }
+        registry.publish_prices(prices_df, model_name=f'{store_mode}_latest', metadata=metadata)
+        logging.info(f"[{store_mode}] Published surge pricing for {len(prices_df)} products")
+
+        return {
+            'n_products': len(prices_df),
+            'registry_status': 'success',
+            'store_mode': store_mode,
+            'mean_demand': float(prices_df['demand_score'].mean()) if 'demand_score' in prices_df.columns else None
+        }
+
+    return {
+        'fetch_interactions': fetch_interactions,
+        'fetch_price_logs': fetch_price_logs,
+        'compute_demand': compute_demand,
+        'aggregate_price_logs': aggregate_price_logs,
+        'join_product_features': join_product_features,
+        'apply_surge_pricing': apply_surge_pricing,
+        'publish_results': publish_results,
+    }
+
+
+def create_surge_pricing_dag(store_mode: str) -> DAG:
+    """Factory: generates a surge pricing DAG for a given store_mode."""
+    callables = _make_task_callables(store_mode)
+
+    dag = DAG(
+        f'surge_pricing_{store_mode}',
+        default_args=DEFAULT_ARGS,
+        description=f'Surge pricing pipeline for {store_mode} store mode',
+        schedule_interval='*/15 * * * *',
+        start_date=days_ago(1),
+        catchup=False,
+        max_active_runs=1,
+        tags=['pricing', 'surge', 'research', store_mode],
+    )
+
+    with dag:
+        t_fetch_interactions = PythonOperator(
+            task_id='fetch_interactions',
+            python_callable=callables['fetch_interactions'],
+            provide_context=True,
+        )
+        t_fetch_price_logs = PythonOperator(
+            task_id='fetch_price_logs',
+            python_callable=callables['fetch_price_logs'],
+            provide_context=True,
+        )
+        t_compute_demand = PythonOperator(
+            task_id='compute_demand',
+            python_callable=callables['compute_demand'],
+            provide_context=True,
+        )
+        t_aggregate_prices = PythonOperator(
+            task_id='aggregate_price_logs',
+            python_callable=callables['aggregate_price_logs'],
+            provide_context=True,
+        )
+        t_join_features = PythonOperator(
+            task_id='join_product_features',
+            python_callable=callables['join_product_features'],
+            provide_context=True,
+        )
+        t_surge_pricing = PythonOperator(
+            task_id='apply_surge_pricing',
+            python_callable=callables['apply_surge_pricing'],
+            provide_context=True,
+        )
+        t_publish = PythonOperator(
+            task_id='publish_results',
+            python_callable=callables['publish_results'],
+            provide_context=True,
+        )
+
+        t_fetch_interactions >> t_compute_demand
+        t_fetch_price_logs >> t_aggregate_prices
+        [t_compute_demand, t_aggregate_prices] >> t_join_features >> t_surge_pricing >> t_publish
+
+    return dag
+
+
+# instantiate DAGs for Airflow to discover
+dag_airline = create_surge_pricing_dag('airline')
+dag_hotel = create_surge_pricing_dag('hotel')
+
+# TODO: Refactor this factory from a surge pricing factory to a general pricing factory
+# We will do this by passing a pricing strategy class to the factory, since the generic pipeline is:
+# take all interaction data, group by sessionId and assign a new price vector to each session
+# in the grouping we get a subset of the interactions per sessionId and we can map that to some Features
+# we define a custom _get_features(interactions .) methodin the strategy class
+# we then run only the inference which is the .predict(trajectory) per-session which will give us a new price vector
+# this we then publish for each sessionId group
+# this might include no deleting most of the pricers we have defined and starting with a super simple surge-pricing algorithm that is no-fit only predict. This we can then test end-to-end and observe changes to prices according to a desired strategy - we have to define this one as a very short term strategy because we run sessions that take only a few minutes.

experiments/airflow/dags/surge_pricing_pipeline.py

@@ -0,0 +1,253 @@
+from airflow import DAG
+from airflow.operators.python import PythonOperator
+from airflow.utils.dates import days_ago
+from datetime import timedelta
+import pandas as pd
+import logging
+import sys
+import pickle
+import io
+
+# add parent dir to path so procesing package can be imported
+sys.path.insert(0, '/opt/airflow')
+
+from procesing.context import PipelineContext
+from procesing.providers import SupabaseProvider, BackendAPIProvider
+from procesing.steps import (
+    FetchInteractionsStep,
+    FetchPriceLogsStep,
+    ComputeDemandStep,
+    AggregatePriceLogsStep,
+    JoinProductFeaturesStep,
+)
+from procesing.pricers.simple import SimpleSurgePricer
+
+default_args = {
+    'owner': 'phantom-research',
+    'depends_on_past': False,
+    'email_on_failure': False,
+    'email_on_retry': False,
+    'retries': 2,
+    'retry_delay': timedelta(minutes=5),
+}
+
+def get_provider():
+    """Factory to create composite provider"""
+    class CompositeProvider(SupabaseProvider, BackendAPIProvider): # TODO: Fix this into one global provider singelton instead of multiple inheritance declarations acoss the codebase
+        def __init__(self):
+            SupabaseProvider.__init__(self)
+            BackendAPIProvider.__init__(self)
+    return CompositeProvider()
+
+def get_context(**kwargs):
+    """Build pipeline context from Airflow config"""
+    dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+    return PipelineContext(
+        provider=get_provider(),
+        store_mode=dag_conf.get('store_mode', 'hotel'),
+    )
+
+# atomic task functions (each wraps one sklearn step)
+def fetch_interactions(**kwargs):
+    """Task: Fetch interaction data from Kafka"""
+    context = get_context(**kwargs)
+    step = FetchInteractionsStep(context)
+    df = step.transform(None)
+
+    kwargs['ti'].xcom_push(key='interactions_raw', value=pickle.dumps(df))
+    logging.info(f"Fetched {len(df)} interaction records")
+    return len(df)
+
+def fetch_price_logs(**kwargs):
+    """Task: Fetch price logs from Kafka"""
+    context = get_context(**kwargs)
+    step = FetchPriceLogsStep(context)
+    df = step.transform(None)
+
+    kwargs['ti'].xcom_push(key='price_logs_raw', value=pickle.dumps(df))
+    logging.info(f"Fetched {len(df)} price records")
+    return len(df)
+
+def compute_demand(**kwargs):
+    """Task: Compute demand scores from interactions"""
+    ti = kwargs['ti']
+    df = pickle.loads(ti.xcom_pull(key='interactions_raw'))
+
+    context = get_context(**kwargs)
+    step = ComputeDemandStep(context)
+    demand_df = step.transform(df)
+    # TODO: clear the xcom
+
+
+    ti.xcom_push(key='demand_data', value=pickle.dumps(demand_df))
+    logging.info(f"Computed demand for {len(demand_df)} products")
+    return len(demand_df)
+
+def aggregate_price_logs(**kwargs):
+    """Task: Aggregate price logs"""
+    ti = kwargs['ti']
+    df = pickle.loads(ti.xcom_pull(key='price_logs_raw'))
+
+    context = get_context(**kwargs)
+    step = AggregatePriceLogsStep(context)
+    price_df = step.transform(df)
+
+    ti.xcom_push(key='price_data', value=pickle.dumps(price_df))
+    logging.info(f"Aggregated price logs for {len(price_df)} products")
+    return len(price_df)
+
+def join_product_features(**kwargs):
+    """Task: Join demand and price data"""
+    ti = kwargs['ti']
+    demand_df = pickle.loads(ti.xcom_pull(key='demand_data'))
+    price_df = pickle.loads(ti.xcom_pull(key='price_data'))
+
+    context = get_context(**kwargs)
+    step = JoinProductFeaturesStep(context)
+    joined_df = step.transform((demand_df, price_df))
+
+    ti.xcom_push(key='product_features', value=pickle.dumps(joined_df))
+    logging.info(f"Joined features for {len(joined_df)} products")
+    return len(joined_df)
+
+def apply_surge_pricing(**kwargs):
+    """Task: Apply surge pricing rules to generate optimal prices"""
+    ti = kwargs['ti']
+    product_features = pickle.loads(ti.xcom_pull(key='product_features'))
+
+    dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+
+    # rename demand_score to demand for pricer compatibility
+    data = product_features.rename(columns={'demand_score': 'demand'})
+
+    high_thresh = dag_conf.get('high_threshold', 10)
+    low_thresh = dag_conf.get('low_threshold', 2)
+    surge_mult = dag_conf.get('surge_multiplier', 1.2)
+    discount_mult = dag_conf.get('discount_multiplier', 0.9)
+
+    logging.info(f"Surge pricing config: high_thresh={high_thresh}, low_thresh={low_thresh}, surge_mult={surge_mult}, discount_mult={discount_mult}")
+    logging.info(f"Demand stats: min={data['demand'].min():.2f}, max={data['demand'].max():.2f}, mean={data['demand'].mean():.2f}")
+    logging.info(f"Products with high demand (>={high_thresh}): {(data['demand'] >= high_thresh).sum()}")
+    logging.info(f"Products with low demand (<={low_thresh}): {(data['demand'] <= low_thresh).sum()}")
+
+    surge_pricer = SimpleSurgePricer(
+        high_threshold=high_thresh,
+        low_threshold=low_thresh,
+        surge_multiplier=surge_mult,
+        discount_multiplier=discount_mult
+    )
+    surge_pricer.fit(data)
+    data['optimal_price'] = surge_pricer.predict()
+
+    base_avg = data['base_price'].mean()
+    optimal_avg = data['optimal_price'].mean()
+    price_change_pct = ((optimal_avg - base_avg) / base_avg) * 100
+
+    logging.info(f"Price adjustment: base_avg={base_avg:.2f}, optimal_avg={optimal_avg:.2f}, change={price_change_pct:+.1f}%")
+
+    prices_df = data[['productId', 'price', 'base_price', 'optimal_price', 'demand']].rename(columns={
+        'price': 'current_price',
+        'demand': 'demand_score'
+    })
+
+    ti.xcom_push(key='predicted_prices', value=pickle.dumps(prices_df))
+    logging.info(f"Applied surge pricing for {len(prices_df)} products")
+    return len(prices_df)
+
+def publish_results(**kwargs):
+    """Task: Publish surge pricing results to registry"""
+    ti = kwargs['ti']
+    prices_df = pickle.loads(ti.xcom_pull(key='predicted_prices'))
+
+    sys.path.insert(0, '/opt/airflow')
+    from lib.model_registry import ModelRegistry
+
+    registry = ModelRegistry()
+    dag_conf = kwargs.get('dag_run').conf if kwargs.get('dag_run') else {}
+
+    metadata = {
+        'timestamp': pd.Timestamp.now().isoformat(),
+        'store_mode': dag_conf.get('store_mode', 'hotel'),
+        'dag_run_id': kwargs['dag_run'].run_id if kwargs.get('dag_run') else 'manual',
+        'pricing_method': 'surge',
+        'high_threshold': dag_conf.get('high_threshold', 10),
+        'low_threshold': dag_conf.get('low_threshold', 2),
+        'surge_multiplier': dag_conf.get('surge_multiplier', 1.2),
+        'discount_multiplier': dag_conf.get('discount_multiplier', 0.9)
+    }
+
+    registry.publish_prices(prices_df, model_name='latest', metadata=metadata)
+
+    logging.info(f"Published surge pricing for {len(prices_df)} products")
+
+    return {
+        'n_products': len(prices_df),
+        'registry_status': 'success',
+        'mean_demand': float(prices_df['demand_score'].mean()) if 'demand_score' in prices_df.columns else None
+    }
+
+
+# DAG definition
+with DAG(
+    'surge_pricing_pipeline',
+    default_args=default_args,
+    description='Simple surge pricing pipeline: demand aggregation + rule-based pricing',
+    schedule_interval='*/15 * * * *',
+    start_date=days_ago(1),
+    catchup=False,
+    max_active_runs=1,
+    tags=['pricing', 'surge', 'research', 'simplified'],
+) as dag:
+
+    # parallel data fetching
+    t_fetch_interactions = PythonOperator(
+        task_id='fetch_interactions',
+        python_callable=fetch_interactions,
+        provide_context=True,
+    )
+
+    t_fetch_price_logs = PythonOperator(
+        task_id='fetch_price_logs',
+        python_callable=fetch_price_logs,
+        provide_context=True,
+    )
+
+    # compute demand from interactions
+    t_compute_demand = PythonOperator(
+        task_id='compute_demand',
+        python_callable=compute_demand,
+        provide_context=True,
+    )
+
+    # aggregate price logs
+    t_aggregate_prices = PythonOperator(
+        task_id='aggregate_price_logs',
+        python_callable=aggregate_price_logs,
+        provide_context=True,
+    )
+
+    # join demand and prices
+    t_join_features = PythonOperator(
+        task_id='join_product_features',
+        python_callable=join_product_features,
+        provide_context=True,
+    )
+
+    # apply surge pricing
+    t_surge_pricing = PythonOperator(
+        task_id='apply_surge_pricing',
+        python_callable=apply_surge_pricing,
+        provide_context=True,
+    )
+
+    # publish to registry
+    t_publish = PythonOperator(
+        task_id='publish_results',
+        python_callable=publish_results,
+        provide_context=True,
+    )
+
+    # dependency graph: parallel fetch -> process -> join -> surge -> publish
+    t_fetch_interactions >> t_compute_demand
+    t_fetch_price_logs >> t_aggregate_prices
+    [t_compute_demand, t_aggregate_prices] >> t_join_features >> t_surge_pricing >> t_publish

experiments/ml/__init__.py

@@ -0,0 +1,21 @@
+from .evals import evaluate
+from .arch import (
+    XGBoostAgentClassifier,
+    LightGBMAgentClassifier,
+    ContrastiveWeakClassifier,
+    TrajectoryEncoder,
+    WeakClassifier,
+    contrastive_loss,
+    featurize_trajectory,
+)
+
+__all__ = [
+    'evaluate',
+    'XGBoostAgentClassifier',
+    'LightGBMAgentClassifier',
+    'ContrastiveWeakClassifier',
+    'TrajectoryEncoder',
+    'WeakClassifier',
+    'contrastive_loss',
+    'featurize_trajectory',
+]

experiments/ml/arch.py

@@ -0,0 +1,212 @@
+# sklearn compatible models for agent detection
+from sklearn.base import BaseEstimator, ClassifierMixin
+from typing import Any, Optional, Tuple, Dict, List
+from abc import ABC, abstractmethod
+from collections import defaultdict
+import numpy as np
+import pandas as pd
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import sys
+from pathlib import Path
+
+# add lib to path for imports
+sys.path.insert(0, str(Path(__file__).parent.parent.parent / 'lib'))
+from lib.features import (
+    transition_histogram as _lib_transition_histogram,
+    temporal_signature as _lib_temporal_signature,
+    state_coverage as _lib_state_coverage,
+    transition_entropy as _lib_transition_entropy,
+    featurize_trajectory as _lib_featurize_trajectory,
+    parse_timestamp
+)
+from lib.state import event_to_state, get_event_name, get_timestamp
+
+TASK = 'classification'
+LABELS = ['human', 'agent']
+
+
+class WeakClassifier(BaseEstimator, ClassifierMixin, ABC):
+    # a simple contrastive machine learning model learns to distinguish human/agent behavior
+    # using weakly supervised contrastive learning + augmentation
+    def __init__(self, **kwargs):
+        super().__init__()
+        self.model = None
+        self.kwargs = kwargs
+
+
+class TrajectoryEncoder(nn.Module):
+    """Encode variable-length event sequences to fixed-dim embedding via bidirectional LSTM"""
+    def __init__(self, input_dim: int, embed_dim: int = 32, hidden_dim: int = 64):
+        super().__init__()
+        self.event_embed = nn.Linear(input_dim, hidden_dim)
+        self.lstm = nn.LSTM(hidden_dim, hidden_dim, batch_first=True, bidirectional=True)
+        self.proj = nn.Linear(hidden_dim * 2, embed_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:  # x: (batch, seq_len, input_dim)
+        h = F.relu(self.event_embed(x))
+        _, (hn, _) = self.lstm(h)
+        hn = torch.cat([hn[-2], hn[-1]], dim=1)  # concat bidirectional hidden states
+        return F.normalize(self.proj(hn), dim=1)  # L2 normalized
+
+
+class ContrastiveWeakClassifier(WeakClassifier):
+    """Contrastive learning classifier for human/agent trajectory discrimination"""
+    def __init__(self, input_dim: int = 64, embed_dim: int = 32, margin: float = 1.0, **kwargs):
+        super().__init__(**kwargs)
+        self.input_dim = input_dim
+        self.embed_dim = embed_dim
+        self.margin = margin
+        self.encoder = TrajectoryEncoder(input_dim, embed_dim)
+        self.classifier = nn.Linear(embed_dim, 2)
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self._fitted = False
+
+    def to_device(self):
+        self.encoder.to(self.device)
+        self.classifier.to(self.device)
+        return self
+
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
+        return self.encoder(x.to(self.device))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        emb = self.encode(x)
+        return self.classifier(emb)
+
+    def fit(self, X, y=None):  # sklearn interface - actual training in weak.train.py
+        self._fitted = True
+        return self
+
+    def predict(self, X: np.ndarray) -> np.ndarray:
+        self.encoder.eval()
+        self.classifier.eval()
+        with torch.no_grad():
+            x = torch.tensor(X, dtype=torch.float32).unsqueeze(1).to(self.device)
+            logits = self.forward(x)
+            return torch.argmax(logits, dim=1).cpu().numpy()
+
+    def predict_proba(self, X: np.ndarray) -> np.ndarray:
+        self.encoder.eval()
+        self.classifier.eval()
+        with torch.no_grad():
+            x = torch.tensor(X, dtype=torch.float32).unsqueeze(1).to(self.device)
+            logits = self.forward(x)
+            return F.softmax(logits, dim=1).cpu().numpy()
+
+
+def contrastive_loss(anchor: torch.Tensor, positive: torch.Tensor, negative: torch.Tensor, margin: float = 0.3) -> torch.Tensor:
+    """Triplet loss using cosine similarity (for L2-normalized embeddings). margin in [0,1] range."""
+    pos_sim = F.cosine_similarity(anchor, positive)  # higher = more similar
+    neg_sim = F.cosine_similarity(anchor, negative)
+    return F.relu(neg_sim - pos_sim + margin).mean()  # want pos_sim > neg_sim + margin
+
+
+def nt_xent_loss(z_i: torch.Tensor, z_j: torch.Tensor, temperature: float = 0.5) -> torch.Tensor:
+    """Normalized temperature-scaled cross entropy loss (SimCLR style)"""
+    batch_size = z_i.size(0)
+    z = torch.cat([z_i, z_j], dim=0)  # (2N, embed_dim)
+    sim = F.cosine_similarity(z.unsqueeze(1), z.unsqueeze(0), dim=2) / temperature
+    mask = torch.eye(2 * batch_size, dtype=torch.bool, device=z.device)
+    sim.masked_fill_(mask, -float('inf'))
+    labels = torch.arange(batch_size, device=z.device)
+    labels = torch.cat([labels + batch_size, labels])  # positive pairs
+    return F.cross_entropy(sim, labels)
+
+
+# feature extraction utilities - delegating to lib.features for unified implementation
+# these wrappers maintain backwards compatibility for existing imports
+
+def transition_histogram(events: List, state_fn, max_states: int = 50) -> np.ndarray:
+    """Compute normalized histogram of state transitions in trajectory"""
+    return _lib_transition_histogram(events, state_fn, max_states)
+
+
+def temporal_signature(events: List, ts_fn) -> np.ndarray:
+    """Extract temporal features: mean/std/skew of inter-event times"""
+    return _lib_temporal_signature(events, ts_fn)
+
+
+def state_coverage(events: List, state_fn, mdp_states: set) -> float:
+    """Fraction of MDP states visited by trajectory"""
+    return _lib_state_coverage(events, state_fn, mdp_states)
+
+
+def transition_entropy(events: List, state_fn) -> float:
+    """Compute entropy of transition distribution (randomness of navigation)"""
+    return _lib_transition_entropy(events, state_fn)
+
+
+def featurize_trajectory(events: List, mdp: Optional[Dict] = None, input_dim: int = 64) -> np.ndarray:
+    """Convert trajectory to fixed-dim feature vector - uses lib.features implementation"""
+    mdp_states = set(mdp.get('states', [])) if mdp else set()
+
+    def _ts_fn(e):
+        return parse_timestamp(get_timestamp(e))
+
+    def _event_name_fn(e):
+        return get_event_name(e)
+
+    return _lib_featurize_trajectory(events, event_to_state, _ts_fn, _event_name_fn, mdp_states, input_dim)
+
+
+# gradient boosting classifiers for comparison baselines
+class XGBoostAgentClassifier(BaseEstimator, ClassifierMixin):
+    """XGBoost classifier for human/agent detection from session features"""
+    def __init__(self, n_estimators: int = 100, max_depth: int = 6, learning_rate: float = 0.1, **kwargs):
+        self.n_estimators = n_estimators
+        self.max_depth = max_depth
+        self.learning_rate = learning_rate
+        self.model = None
+        self.kwargs = kwargs
+
+    def fit(self, X: np.ndarray, y: np.ndarray):
+        try:
+            import xgboost as xgb
+            self.model = xgb.XGBClassifier(n_estimators=self.n_estimators, max_depth=self.max_depth,
+                                           learning_rate=self.learning_rate, **self.kwargs)
+            self.model.fit(X, y)
+        except ImportError:
+            raise ImportError("xgboost required for XGBoostAgentClassifier")
+        return self
+
+    def predict(self, X: np.ndarray) -> np.ndarray:
+        if self.model is None:
+            raise ValueError("fit the model first")
+        return self.model.predict(X)
+
+    def predict_proba(self, X: np.ndarray) -> np.ndarray:
+        if self.model is None:
+            raise ValueError("fit the model first")
+        return self.model.predict_proba(X)
+
+
+class LightGBMAgentClassifier(BaseEstimator, ClassifierMixin):
+    """LightGBM classifier for human/agent detection from session features"""
+    def __init__(self, n_estimators: int = 100, max_depth: int = -1, learning_rate: float = 0.1, **kwargs):
+        self.n_estimators = n_estimators
+        self.max_depth = max_depth
+        self.learning_rate = learning_rate
+        self.model = None
+        self.kwargs = kwargs
+
+    def fit(self, X: np.ndarray, y: np.ndarray):
+        try:
+            import lightgbm as lgb
+            self.model = lgb.LGBMClassifier(n_estimators=self.n_estimators, max_depth=self.max_depth,
+                                            learning_rate=self.learning_rate, verbose=-1, **self.kwargs)
+            self.model.fit(X, y)
+        except ImportError:
+            raise ImportError("lightgbm required for LightGBMAgentClassifier")
+        return self
+
+    def predict(self, X: np.ndarray) -> np.ndarray:
+        if self.model is None:
+            raise ValueError("fit the model first")
+        return self.model.predict(X)
+
+    def predict_proba(self, X: np.ndarray) -> np.ndarray:
+        if self.model is None:
+            raise ValueError("fit the model first")
+        return self.model.predict_proba(X)

experiments/ml/encoder/__init__.py

@@ -0,0 +1 @@
+from .encoder import Window, extract_windows, build_windows, WindowDataset, PrototypeClassifier, train, loocv

experiments/ml/encoder/encoder.py

@@ -0,0 +1,210 @@
+"""Contrastive encoder via trajectory windowing. Classification by prototype distance."""
+import sys
+sys.path.insert(0, "/home/velocitatem/Documents/Projects/PHANTOM/sim/rl/behavior_loader")
+sys.path.insert(0, "/home/velocitatem/Documents/Projects/PHANTOM/experiments/ml")
+
+from sim.rl.behavior_loader.loader import JointLoader, PayloadModel
+from arch import TrajectoryEncoder, featurize_trajectory, nt_xent_loss
+from typing import List, Dict, Tuple
+from dataclasses import dataclass
+from datetime import datetime
+import numpy as np, torch, torch.nn.functional as F, random, optuna
+from torch.utils.data import Dataset, DataLoader
+from torch.optim import Adam
+from torch.utils.tensorboard import SummaryWriter
+
+RUNS = "/home/velocitatem/Documents/Projects/PHANTOM/experiments/ml/runs"
+AGENT_DIR = "/home/velocitatem/Documents/Projects/PHANTOM/experiments/agents/collected_data/"
+HUMAN_DIR = "/home/velocitatem/Documents/Projects/PHANTOM/experiments/collected_data/"
+
+
+@dataclass
+class Window:
+    events: List[PayloadModel]
+    traj_id: str
+    label: int  # 0=human, 1=agent
+
+
+def extract_windows(events: List[PayloadModel], traj_id: str, label: int,
+                    sizes: List[int] = [5, 10, 15], stride: int = 2) -> List[Window]:
+    """Multi-scale overlapping windows from trajectory"""
+    n = len(events)
+    wins = [Window(events[i:i+s], traj_id, label) for s in sizes if n >= s for i in range(0, n-s+1, stride)]
+    if n >= 3: wins.append(Window(events, traj_id, label))  # full traj
+    return wins
+
+
+def build_windows(data: Dict[str, List], sizes=[5,10,15], stride=2) -> List[Window]:
+    return [w for tid, evts in data.items()
+            for w in extract_windows(evts, tid, 0 if tid.startswith('human_') else 1, sizes, stride)]
+
+
+class WindowDataset(Dataset):
+    """Yields (anchor, positive) pairs from same class"""
+    def __init__(self, windows: List[Window], dim: int = 64):
+        self.wins, self.dim = windows, dim
+        self.by_label = {0: [i for i,w in enumerate(windows) if w.label==0],
+                         1: [i for i,w in enumerate(windows) if w.label==1]}
+        self.by_traj = {}
+        for i, w in enumerate(windows): self.by_traj.setdefault(w.traj_id, []).append(i)
+
+    def __len__(self): return len(self.wins)
+
+    def _feat(self, evts): return featurize_trajectory(evts, None, self.dim)
+
+    def _aug(self, evts):  # subsample 70-100%
+        if len(evts) < 4: return evts
+        k = max(3, int(len(evts) * random.uniform(0.7, 1.0)))
+        start = random.randint(0, len(evts) - k)
+        return evts[start:start+k]
+
+    def __getitem__(self, idx):
+        w = self.wins[idx]
+        pool = [i for i in self.by_label[w.label] if self.wins[i].traj_id != w.traj_id]
+        pos_idx = random.choice(pool) if pool else idx
+        a = torch.tensor(self._feat(self._aug(w.events)), dtype=torch.float32)
+        p = torch.tensor(self._feat(self._aug(self.wins[pos_idx].events)), dtype=torch.float32)
+        return a, p, w.label
+
+
+class PrototypeClassifier:
+    """Classify by distance to class centroids"""
+    def __init__(self, encoder: TrajectoryEncoder, device = 'cuda', dim=64):
+        self.enc, self.dev, self.dim = encoder, device, dim
+        self.centroids = {0: None, 1: None}
+
+    def fit(self, windows: List[Window]):
+        self.enc.eval()
+        embs = {0: [], 1: []}
+        with torch.no_grad():
+            for w in windows:
+                x = torch.tensor(featurize_trajectory(w.events, None, self.dim), dtype=torch.float32)
+                z = self.enc(x.unsqueeze(0).unsqueeze(1).to(self.dev))
+                embs[w.label].append(z)
+        self.centroids = {k: torch.cat(v).mean(0, keepdim=True) if v else None for k, v in embs.items()}
+        return self
+
+    def predict(self, events: List[PayloadModel]) -> Tuple[int, float, Dict]:
+        """Returns (pred, confidence, debug). Confidence via softmax over -distances."""
+        self.enc.eval()
+        with torch.no_grad():
+            x = torch.tensor(featurize_trajectory(events, None, self.dim), dtype=torch.float32)
+            z = self.enc(x.unsqueeze(0).unsqueeze(1).to(self.dev))
+            dists = {k: torch.norm(z - c, dim=1).item() for k, c in self.centroids.items() if c is not None}
+            if not dists: return 0, 0.0, {'d': {}, 'p': [0.5, 0.5]}
+            pred = min(dists, key=dists.get)
+            d0, d1 = dists.get(0, 1e6), dists.get(1, 1e6)  # softmax(-d) gives higher prob to closer centroid
+            probs = F.softmax(torch.tensor([[-d0, -d1]]), dim=1).squeeze()
+        return pred, probs[pred].item(), {'d': dists, 'p': probs.tolist()}
+
+
+def train(epochs=200, lr=5e-4, batch=16, dim=64, emb=32, temp=0.5,
+          sizes=[5,10,15], stride=2, name=None, verbose=True):
+    data = JointLoader(HUMAN_DIR, AGENT_DIR).get_data()
+    wins = build_windows(data, sizes, stride)
+    if verbose: print(f"Windows: {len(wins)} ({sum(w.label==0 for w in wins)}h/{sum(w.label==1 for w in wins)}a)")
+
+    dev = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    enc = TrajectoryEncoder(dim, emb).to(dev)
+    opt = Adam(enc.parameters(), lr=lr)
+    loader = DataLoader(WindowDataset(wins, dim), batch_size=batch, shuffle=True, drop_last=True)
+
+    name = name or f"enc_{dim}_{emb}_{datetime.now():%Y%m%d_%H%M%S}"
+    writer = SummaryWriter(f"{RUNS}/encoder/{name}")
+
+    for ep in range(epochs):
+        enc.train()
+        total, n = 0.0, 0
+        for a, p, _ in loader:
+            loss = nt_xent_loss(enc(a.unsqueeze(1).to(dev)), enc(p.unsqueeze(1).to(dev)), temp)
+            opt.zero_grad(); loss.backward(); opt.step()
+            total += loss.item(); n += 1
+        avg = total / max(n, 1)
+        writer.add_scalar('loss-ntxent', avg, ep)
+        if verbose and (ep+1) % 20 == 0: print(f"Epoch {ep+1}: {avg:.4f}")
+
+    writer.close()
+    return enc, wins, dev
+
+
+def loocv(epochs=100, lr=5e-4, dim=64, emb=32, temp=0.5, sizes=[5,10,15], stride=2, verbose=True):
+    """Leave-one-trajectory-out CV"""
+    data = JointLoader(HUMAN_DIR, AGENT_DIR).get_data()
+    dev = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    results = []
+
+    for test_id in data:
+        train_data = {k: v for k, v in data.items() if k != test_id}
+        if not any(k.startswith('human_') for k in train_data) or not any(k.startswith('agent_') for k in train_data):
+            continue
+
+        wins = build_windows(train_data, sizes, stride)
+        enc = TrajectoryEncoder(dim, emb).to(dev)
+        opt = Adam(enc.parameters(), lr=lr)
+        loader = DataLoader(WindowDataset(wins, dim), batch_size=min(16, len(wins)//2 or 1),
+                            shuffle=True, drop_last=len(wins)>2)
+
+        for _ in range(epochs):
+            enc.train()
+            for a, p, _ in loader:
+                loss = nt_xent_loss(enc(a.unsqueeze(1).to(dev)), enc(p.unsqueeze(1).to(dev)), temp)
+                opt.zero_grad(); loss.backward(); opt.step()
+
+        clf = PrototypeClassifier(enc, dev, dim).fit(wins)
+        pred, conf, dbg = clf.predict(data[test_id])
+        actual = 0 if test_id.startswith('human_') else 1
+        results.append((pred, actual, conf))
+        if verbose: print(f"{test_id[:18]}: pred={pred} conf={conf:.2f} actual={actual} {'OK' if pred==actual else 'MISS'}")
+
+    if results:
+        acc = sum(p==a for p,a,_ in results) / len(results)
+        if verbose: print(f"\nAccuracy: {acc:.1%} ({sum(p==a for p,a,_ in results)}/{len(results)})")
+        return acc, results
+    return 0.0, []
+
+
+def hparam_tune(n_trials=50, epochs=60, n_jobs=2, verbose=True):
+    """Optuna hyperparameter search maximizing LOOCV accuracy"""
+    def objective(trial):
+        lr = trial.suggest_float('lr', 1e-5, 1e-2, log=True)
+        dim = trial.suggest_categorical('dim', [32, 64, 128, 256])
+        emb = trial.suggest_categorical('emb', [16, 32, 64, 128])
+        temp = trial.suggest_float('temp', 0.05, 1.0)
+        stride = trial.suggest_int('stride', 1, 4)
+        sizes = [trial.suggest_int(f's{i}', 3, 20) for i in range(3)]
+        sizes = sorted(set(sizes))  # unique sorted
+        acc, _ = loocv(epochs, lr, dim, emb, temp, sizes, stride, verbose=False)
+        return acc
+
+    study = optuna.create_study(direction='maximize', study_name='encoder_hparam',
+                                sampler=optuna.samplers.TPESampler(seed=42))
+    study.optimize(objective, n_trials=n_trials, n_jobs=n_jobs, show_progress_bar=verbose)
+
+    best = study.best_params
+    if verbose:
+        print(f"\nBest accuracy: {study.best_value:.1%}")
+        print(f"Best params: {best}")
+    return best, study
+
+
+if __name__ == "__main__":
+    import argparse
+    p = argparse.ArgumentParser()
+    p.add_argument('--mode', choices=['train', 'eval', 'hparam'], default='train')
+    p.add_argument('--epochs', type=int, default=200)
+    p.add_argument('--lr', type=float, default=5e-4)
+    p.add_argument('--dim', type=int, default=128)
+    p.add_argument('--emb', type=int, default=64)
+    p.add_argument('--temp', type=float, default=0.1)
+    p.add_argument('--sizes', type=str, default='5,10,15')
+    p.add_argument('--stride', type=int, default=2)
+    p.add_argument('--n_trials', type=int, default=50)
+    args = p.parse_args()
+    sizes = [int(x) for x in args.sizes.split(',')]
+
+    if args.mode == 'train':
+        enc, wins, dev = train(args.epochs, args.lr, 16, args.dim, args.emb, args.temp, sizes, args.stride)
+    elif args.mode == 'hparam':
+        best, study = hparam_tune(args.n_trials, min(args.epochs, 60))
+    else:
+        loocv(args.epochs, args.lr, args.dim, args.emb, args.temp, sizes, args.stride)

experiments/ml/evals.py

@@ -0,0 +1,103 @@
+from sklearn.metrics import (accuracy_score, precision_score, recall_score,
+                             f1_score, roc_auc_score, confusion_matrix, roc_curve)
+from torch.utils.tensorboard import SummaryWriter
+from logging import getLogger
+import numpy as np
+import matplotlib.pyplot as plt
+import io
+from PIL import Image
+
+logger = getLogger(__name__)
+
+
+def log_feature_importance(writer, model, feature_names, epoch):
+    """Visualize and log feature importance to TensorBoard"""
+    if not hasattr(model, 'feature_importances_') or model.feature_importances_ is None:
+        return
+
+    importance = model.feature_importances_
+    indices = np.argsort(importance)[::-1][:20]  # top 20
+    top_features = [feature_names[i] for i in indices]
+    top_importance = importance[indices]
+
+    for i, (feat, imp) in enumerate(zip(top_features, top_importance)):
+        writer.add_scalar(f'FeatureImportance/{feat}', imp, epoch)
+
+    fig, ax = plt.subplots(figsize=(10, 8))
+    ax.barh(range(len(top_features)), top_importance, align='center')
+    ax.set_yticks(range(len(top_features)))
+    ax.set_yticklabels(top_features)
+    ax.invert_yaxis()
+    ax.set_xlabel('Importance')
+    ax.set_title(f'Top 20 Feature Importance (Epoch {epoch})')
+    ax.grid(axis='x', alpha=0.3)
+
+    buf = io.BytesIO()
+    plt.tight_layout()
+    plt.savefig(buf, format='png', dpi=100)
+    buf.seek(0)
+    img = Image.open(buf)
+    img_arr = np.array(img)
+    writer.add_image('FeatureImportance/Chart', img_arr, epoch, dataformats='HWC')
+    plt.close()
+
+def evaluate(perdicted_class, predicted_proba, true_class, writer: SummaryWriter, epoch: int):
+    accuracy = accuracy_score(true_class, perdicted_class)
+    precision = precision_score(true_class, perdicted_class, zero_division=0)
+    recall = recall_score(true_class, perdicted_class, zero_division=0)
+    f1 = f1_score(true_class, perdicted_class, zero_division=0)
+    roc_auc = roc_auc_score(true_class, predicted_proba)
+
+    writer.add_scalar('Eval/Accuracy', accuracy, epoch)
+    writer.add_scalar('Eval/Precision', precision, epoch)
+    writer.add_scalar('Eval/Recall', recall, epoch)
+    writer.add_scalar('Eval/F1_Score', f1, epoch)
+    writer.add_scalar('Eval/ROC_AUC', roc_auc, epoch)
+
+    # confusion matrix
+    cm = confusion_matrix(true_class, perdicted_class)
+    tn, fp, fn, tp = cm.ravel()
+    writer.add_scalar('Eval/TrueNeg', tn, epoch)
+    writer.add_scalar('Eval/FalsePos', fp, epoch)
+    writer.add_scalar('Eval/FalseNeg', fn, epoch)
+    writer.add_scalar('Eval/TruePos', tp, epoch)
+
+    # specificity and sensitivity
+    specificity = tn / (tn + fp) if (tn + fp) > 0 else 0
+    sensitivity = recall  # same as recall/TPR
+    writer.add_scalar('Eval/Specificity', specificity, epoch)
+    writer.add_scalar('Eval/Sensitivity', sensitivity, epoch)
+
+    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))
+    ax1.matshow(cm, cmap='Blues', alpha=0.7)
+    for i in range(2):
+        for j in range(2):
+            ax1.text(j, i, str(cm[i, j]), ha='center', va='center', fontsize=14)
+    ax1.set_xlabel('Predicted')
+    ax1.set_ylabel('True')
+    ax1.set_title(f'Confusion Matrix (Epoch {epoch})')
+    ax1.set_xticks([0, 1])
+    ax1.set_yticks([0, 1])
+    ax1.set_xticklabels(['Human', 'Agent'])
+    ax1.set_yticklabels(['Human', 'Agent'])
+
+    # ROC curve
+    fpr, tpr, _ = roc_curve(true_class, predicted_proba)
+    ax2.plot(fpr, tpr, label=f'AUC={roc_auc:.3f}', linewidth=2)
+    ax2.plot([0, 1], [0, 1], 'k--', label='Random')
+    ax2.set_xlabel('False Positive Rate')
+    ax2.set_ylabel('True Positive Rate')
+    ax2.set_title('ROC Curve')
+    ax2.legend()
+    ax2.grid(alpha=0.3)
+
+    buf = io.BytesIO()
+    plt.tight_layout()
+    plt.savefig(buf, format='png', dpi=100)
+    buf.seek(0)
+    img = Image.open(buf)
+    img_arr = np.array(img)
+    writer.add_image('Eval/Metrics', img_arr, epoch, dataformats='HWC')
+    plt.close()
+
+    logger.info(f"Eval {epoch}: Acc={accuracy:.4f} Prec={precision:.4f} Rec={recall:.4f} F1={f1:.4f} AUC={roc_auc:.4f}")

experiments/ml/requirements.txt

@@ -0,0 +1,6 @@
+torch
+tensorboard
+fastparquet
+pyarrow
+xgboost
+lightgbm

experiments/ml/train.py

@@ -0,0 +1,137 @@
+from torch.utils.tensorboard import SummaryWriter
+from sklearn.model_selection import train_test_split
+from logging import getLogger
+from pathlib import Path
+import pandas as pd
+import numpy as np
+import joblib
+from datetime import datetime
+from ml.evals import evaluate, log_feature_importance
+from ml.arch import XGBoostAgentClassifier, LightGBMAgentClassifier, LABELS
+
+logger = getLogger(__name__)
+
+FEATURE_COLS_EXCLUDE = ['sessionId', 'experimentId', 'is_agent', 'xp_human_only', 'xp_market_mode', 'browser_family']
+RUNS_DIR = Path('ml/runs')
+CHECKPOINTS_DIR = Path('ml/checkpoints')
+
+
+def prepare_data(df):
+    """
+    Prepare feature matrix and labels from raw dataframe
+    Handles missing labels, feature selection, and categorical encoding
+    Returns: (X, y, feature_cols)
+    """
+    # drop rows with missing labels
+    n_before = len(df)
+    df = df[df['is_agent'].notna()].copy()
+    n_dropped = n_before - len(df)
+    if n_dropped > 0:
+        logger.warning(f"Dropped {n_dropped} sessions with missing labels")
+
+    if len(df) == 0:
+        logger.error("No labeled data available")
+        return None, None, None
+
+    feature_cols = [c for c in df.columns if c not in FEATURE_COLS_EXCLUDE]
+
+    # handle categorical browser_family via one-hot encoding
+    if 'browser_family' in df.columns:
+        browser_dummies = pd.get_dummies(df['browser_family'], prefix='browser', drop_first=True)
+        df = pd.concat([df, browser_dummies], axis=1)
+        feature_cols.extend(browser_dummies.columns.tolist())
+
+    X = df[feature_cols].fillna(0)
+    y = df['is_agent'].astype(int)
+
+    return X, y, feature_cols
+
+
+def train(data_path=None, model_type='xgboost', test_size=0.2, random_state=42,
+          n_estimators=200, max_depth=6, learning_rate=0.05):
+    """
+    Train agent detection classifier
+    Args:
+        data_path: path to labeled feature matrix CSV or parquet
+        model_type: 'xgboost' or 'lightgbm'
+        test_size: fraction for test split
+        random_state: seed for reproducibility
+    """
+    RUNS_DIR.mkdir(exist_ok=True)
+    CHECKPOINTS_DIR.mkdir(exist_ok=True)
+
+    run_name = f"{model_type}_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
+    writer = SummaryWriter(log_dir=RUNS_DIR / run_name)
+    logger.info(f"Starting training run: {run_name}")
+
+    # load data
+    if data_path is None:
+        logger.error("data_path required")
+        return
+    df = pd.read_parquet(data_path)
+    logger.info(f"Loaded {len(df)} sessions from {data_path}")
+
+    # prepare features and labels
+    if 'is_agent' not in df.columns:
+        logger.error("Missing is_agent column")
+        return
+
+    X, y, feature_cols = prepare_data(df)
+    if X is None:
+        return
+
+    # class distribution
+    n_agents = y.sum()
+    n_humans = (y == 0).sum()
+    logger.info(f"Class distribution: {n_humans} humans, {n_agents} agents" + (f" (ratio {n_humans / n_agents:.2f})" if n_agents > 0 else ""))
+
+    # train/test split with stratification
+    X_train, X_test, y_train, y_test = train_test_split(
+        X, y, test_size=test_size, random_state=random_state, stratify=y
+    )
+    logger.info(f"Train: {len(X_train)}, Test: {len(X_test)}")
+
+    # init model
+    if model_type == 'xgboost':
+        model = XGBoostAgentClassifier(
+            n_estimators=n_estimators,
+            max_depth=max_depth,
+            learning_rate=learning_rate
+        )
+    elif model_type == 'lightgbm':
+        model = LightGBMAgentClassifier(
+            n_estimators=n_estimators,
+            max_depth=max_depth,
+            learning_rate=learning_rate
+        )
+    else:
+        logger.error(f"Unknown model type: {model_type}")
+        return
+
+    # train with eval set for early stopping
+    model.fit(X_train, y_train, eval_set=[(X_test, y_test)])
+    logger.info("Training complete")
+
+    # evaluate on test set
+    y_pred = model.predict(X_test)
+    y_prob = model.predict_proba(X_test)[:, 1]
+
+    evaluate(y_pred, y_prob, y_test, writer, epoch=0)
+
+    # log feature importance
+    log_feature_importance(writer, model, X.columns.tolist(), epoch=0)
+
+    # save model
+    model_path = CHECKPOINTS_DIR / f"{run_name}.pkl"
+    joblib.dump({'model': model, 'feature_cols': X.columns.tolist(), 'run_name': run_name}, model_path)
+    logger.info(f"Model saved to {model_path}")
+
+    writer.close()
+    return model, X.columns.tolist()
+
+
+if __name__ == "__main__":
+    import sys
+    data_path = sys.argv[1]
+    model_type = sys.argv[2] if len(sys.argv) > 2 else 'xgboost'
+    train(data_path, model_type=model_type)

experiments/ml/weak_train.py

@@ -0,0 +1,246 @@
+import sys
+sys.path.insert(0, "/home/velocitatem/Documents/Projects/PHANTOM/sim/rl/behavior_loader")
+sys.path.insert(0, "/home/velocitatem/Documents/Projects/PHANTOM/experiments/ml")
+
+from sim.rl.behavior_loader.loader import AgentLoader, Loader, JointLoader, PayloadModel
+from sim.rl.behavior_loader.models import JointBehaviorModel
+from arch import ContrastiveWeakClassifier, contrastive_loss, featurize_trajectory
+from typing import List, Optional, Dict
+from datetime import datetime, timedelta
+from copy import deepcopy
+import numpy as np
+import random
+import torch
+from torch.utils.data import Dataset, DataLoader
+from torch.optim import Adam
+from torch.utils.tensorboard import SummaryWriter
+
+RUNS_DIR = "/home/velocitatem/Documents/Projects/PHANTOM/experiments/ml/runs"
+agent_dir = "/home/velocitatem/Documents/Projects/PHANTOM/experiments/agents/collected_data/"
+human_dir = "/home/velocitatem/Documents/Projects/PHANTOM/experiments/collected_data/"
+
+
+def _perturb_ts(evt: PayloadModel, jitter_ms: int = 500) -> PayloadModel:
+    """Add random jitter to event timestamp"""
+    new_evt = deepcopy(evt)
+    try:
+        ts = datetime.fromisoformat(evt.ts.replace('Z', '+00:00'))
+        delta = timedelta(milliseconds=random.randint(-jitter_ms, jitter_ms))
+        new_evt.ts = (ts + delta).isoformat()
+    except:
+        pass
+    return new_evt
+
+
+def augment_trajectory(trajectory: List[PayloadModel], rate: float = 0.1) -> List[PayloadModel]:
+    """Apply random augmentation to trajectory for contrastive learning"""
+    if len(trajectory) < 2:
+        return trajectory
+
+    aug_type = random.choice(['window', 'shuffle', 'noise', 'drop'])
+
+    if aug_type == 'window':  # random contiguous sub-sequence (70-100% length)
+        min_len = max(2, int(len(trajectory) * 0.7))
+        sub_len = random.randint(min_len, len(trajectory))
+        start = random.randint(0, len(trajectory) - sub_len)
+        return trajectory[start:start + sub_len]
+
+    elif aug_type == 'shuffle':  # swap adjacent pairs with probability rate
+        result = list(trajectory)
+        for i in range(len(result) - 1):
+            if random.random() < rate:
+                result[i], result[i + 1] = result[i + 1], result[i]
+        return result
+
+    elif aug_type == 'drop':  # drop events with probability rate
+        result = [e for e in trajectory if random.random() > rate]
+        return result if len(result) >= 2 else trajectory[:2]
+
+    elif aug_type == 'noise':  # perturb timestamps
+        return [_perturb_ts(e, jitter_ms=500) for e in trajectory]
+
+    return trajectory
+
+
+class TripletDataset(Dataset):
+    """Generate (anchor, positive, negative) triplets on-the-fly with augmentation"""
+    def __init__(self, data: Dict[str, List[PayloadModel]], mdp: Optional[Dict], augment_fn, input_dim: int = 64, multiplier: int = 10):
+        self.sessions = list(data.items())
+        self.human_ids = [i for i, (sid, _) in enumerate(self.sessions) if sid.startswith('human_')]
+        self.agent_ids = [i for i, (sid, _) in enumerate(self.sessions) if sid.startswith('agent_')]
+        self.mdp = mdp
+        self.augment = augment_fn
+        self.input_dim = input_dim
+        self.multiplier = multiplier
+
+        if not self.human_ids or not self.agent_ids:
+            raise ValueError(f"Need both human ({len(self.human_ids)}) and agent ({len(self.agent_ids)}) sessions")
+
+    def __len__(self) -> int:
+        return len(self.sessions) * self.multiplier
+
+    def __getitem__(self, idx: int):
+        anchor_idx = idx % len(self.sessions)
+        sid, events = self.sessions[anchor_idx]
+        is_human = sid.startswith('human_')
+
+        anchor = featurize_trajectory(events, self.mdp, self.input_dim)
+        positive = featurize_trajectory(self.augment(events), self.mdp, self.input_dim)
+
+        neg_pool = self.agent_ids if is_human else self.human_ids
+        neg_idx = random.choice(neg_pool)
+        negative = featurize_trajectory(self.sessions[neg_idx][1], self.mdp, self.input_dim)
+
+        label = 0 if is_human else 1  # 0=human, 1=agent
+        return (torch.tensor(anchor, dtype=torch.float32),
+                torch.tensor(positive, dtype=torch.float32),
+                torch.tensor(negative, dtype=torch.float32),
+                torch.tensor(label, dtype=torch.long))
+
+
+def train(epochs: int = 100, lr: float = 1e-3, batch_size: int = 4, input_dim: int = 64,
+          embed_dim: int = 32, margin: float = 0.3, verbose: bool = True, run_name: str = None):
+    """Train contrastive weak classifier on human/agent trajectories"""
+    joint = JointLoader(human_dir, agent_dir)
+    data = joint.get_data()
+    if verbose:
+        print(f"Loaded {len(data)} sessions")
+
+    joint_model = JointBehaviorModel(human_dir, agent_dir)
+    ref_mdp = joint_model.build_MDP()
+
+    dataset = TripletDataset(data, ref_mdp, augment_trajectory, input_dim=input_dim)
+    loader = DataLoader(dataset, batch_size=batch_size, shuffle=True, drop_last=True)
+
+    model = ContrastiveWeakClassifier(input_dim=input_dim, embed_dim=embed_dim, margin=margin)
+    model.to_device()
+
+    run_name = run_name or f"d{input_dim}_e{embed_dim}_lr{lr}_m{margin}_{datetime.now():%Y%m%d_%H%M%S}"
+    writer = SummaryWriter(f"{RUNS_DIR}/train/{run_name}")
+
+    optimizer = Adam(list(model.encoder.parameters()) + list(model.classifier.parameters()), lr=lr)
+    ce_loss_fn = torch.nn.CrossEntropyLoss()
+
+    best_loss = float('inf')
+    for epoch in range(epochs):
+        model.encoder.train()
+        model.classifier.train()
+        total_loss, n_batches = 0.0, 0
+
+        for anchor, positive, negative, labels in loader:
+            anchor, positive, negative, labels = [t.to(model.device) for t in [anchor, positive, negative, labels]]
+            z_a, z_p, z_n = [model.encoder(t.unsqueeze(1)) for t in [anchor, positive, negative]]
+
+            trip_loss = contrastive_loss(z_a, z_p, z_n, margin=model.margin)
+            ce = ce_loss_fn(model.classifier(z_a), labels)
+            loss = trip_loss + 0.5 * ce
+
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+            total_loss += loss.item()
+            n_batches += 1
+
+        avg_loss = total_loss / max(n_batches, 1)
+        writer.add_scalar('loss', avg_loss, epoch)
+
+        if verbose and (epoch + 1) % 10 == 0:
+            print(f"Epoch {epoch+1}/{epochs}: loss={avg_loss:.4f}")
+        if avg_loss < best_loss:
+            best_loss = avg_loss
+
+    writer.close()
+    if verbose:
+        print(f"Done. Best={best_loss:.4f} TB:{RUNS_DIR}/train/{run_name}")
+
+    return model, ref_mdp
+
+
+def evaluate_loocv(input_dim: int = 64, embed_dim: int = 32, epochs_per_fold: int = 50,
+                   lr: float = 1e-3, margin: float = 0.3, run_name: str = None):
+    """Leave-one-out cross-validation given limited samples"""
+    joint = JointLoader(human_dir, agent_dir)
+    data = joint.get_data()
+    session_ids = list(data.keys())
+
+    joint_model = JointBehaviorModel(human_dir, agent_dir)
+    ref_mdp = joint_model.build_MDP()
+
+    run_name = run_name or f"loocv_d{input_dim}_e{embed_dim}_m{margin}_{datetime.now():%Y%m%d_%H%M%S}"
+    writer = SummaryWriter(f"{RUNS_DIR}/eval/{run_name}")
+
+    predictions, actuals = [], []
+
+    for fold_idx, test_sid in enumerate(session_ids):
+        train_data = {k: v for k, v in data.items() if k != test_sid}
+        test_events = data[test_sid]
+        test_label = 0 if test_sid.startswith('human_') else 1
+
+        n_human = sum(1 for k in train_data if k.startswith('human_'))
+        n_agent = sum(1 for k in train_data if k.startswith('agent_'))
+        if n_human == 0 or n_agent == 0:
+            continue
+
+        try:
+            dataset = TripletDataset(train_data, ref_mdp, augment_trajectory, input_dim=input_dim, multiplier=5)
+            loader = DataLoader(dataset, batch_size=2, shuffle=True, drop_last=True)
+
+            model = ContrastiveWeakClassifier(input_dim=input_dim, embed_dim=embed_dim, margin=margin)
+            model.to_device()
+            optimizer = Adam(list(model.encoder.parameters()) + list(model.classifier.parameters()), lr=lr)
+
+            model.encoder.train()
+            model.classifier.train()
+            for _ in range(epochs_per_fold):
+                for anchor, positive, negative, labels in loader:
+                    z_a, z_p, z_n = [model.encoder(t.unsqueeze(1).to(model.device)) for t in [anchor, positive, negative]]
+                    loss = contrastive_loss(z_a, z_p, z_n, margin=margin)
+                    optimizer.zero_grad()
+                    loss.backward()
+                    optimizer.step()
+
+            test_feat = featurize_trajectory(test_events, ref_mdp, input_dim)
+            pred = model.predict(test_feat.reshape(1, -1))[0]
+            predictions.append(pred)
+            actuals.append(test_label)
+            print(f"  {test_sid[:12]}...: pred={pred}, actual={test_label}, {'OK' if pred == test_label else 'MISS'}")
+
+        except Exception as e:
+            print(f"Error: {e}")
+
+    if predictions:
+        acc = sum(p == a for p, a in zip(predictions, actuals)) / len(predictions)
+        tp = sum(1 for p, a in zip(predictions, actuals) if p == 1 and a == 1)
+        fp = sum(1 for p, a in zip(predictions, actuals) if p == 1 and a == 0)
+        fn = sum(1 for p, a in zip(predictions, actuals) if p == 0 and a == 1)
+        prec, rec = tp / max(tp + fp, 1), tp / max(tp + fn, 1)
+        f1 = 2 * prec * rec / max(prec + rec, 1e-10)
+        writer.add_scalar('accuracy', acc, 0)
+        writer.add_scalar('f1', f1, 0)
+        writer.add_scalar('precision', prec, 0)
+        writer.add_scalar('recall', rec, 0)
+        writer.close()
+        print(f"\nAccuracy: {acc:.2%} F1: {f1:.3f} TB:{RUNS_DIR}/eval/{run_name}")
+        return acc, predictions, actuals
+    writer.close()
+    return 0.0, [], []
+
+
+if __name__ == "__main__":
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--mode', choices=['train', 'eval'], default='train')
+    parser.add_argument('--epochs', type=int, default=100)
+    parser.add_argument('--lr', type=float, default=1e-3)
+    parser.add_argument('--margin', type=float, default=0.3)
+    parser.add_argument('--input-dim', type=int, default=64)
+    parser.add_argument('--embed-dim', type=int, default=32)
+    parser.add_argument('--run-name', type=str, default=None)
+    args = parser.parse_args()
+
+    if args.mode == 'train':
+        model, mdp = train(epochs=args.epochs, lr=args.lr, input_dim=args.input_dim,
+                           embed_dim=args.embed_dim, margin=args.margin, run_name=args.run_name)
+    else:
+        evaluate_loocv(input_dim=args.input_dim, embed_dim=args.embed_dim, epochs_per_fold=args.epochs,
+                       lr=args.lr, margin=args.margin, run_name=args.run_name)