19 KiB
PHANTOM: setup for operators and partners
This guide walks a team from business context (what you sell, how you price, what traffic you worry about) through a running PHANTOM stack, behavioral kernels and contamination, and RL training / benchmarking. The math lives in the thesis PDF; here we tie operations to that math without re-deriving it. References to the thesis use chapter numbers only (build the PDF locally if you need line-level citations).
Thesis (PDF): thesis-latest.pdf
1. Who this is for / prerequisites
Audience: Engineers and researchers who run Docker, a Next.js app, and Python tooling; product or risk stakeholders who define experiment goals and acceptable UX tradeoffs.
Skills: Docker Compose, Node/npm, Python 3.8+, basic Kafka/Redis mental model.
Decide up front:
- Vertical vs demo: The repo ships
hotelandairlinestorefront modes (STORE_MODE). Anything beyond that is custom integration work. - Data residency: Event streams and training artifacts default to paths under the repo (overridable via
PHANTOM_* env vars inlib/config.py). Decide where logs and models may live before you point production-like traffic at the stack. - Experiment governance: Who may run human vs agent sessions, how sessions are labeled or weak-labeled for research, and retention policy for interaction logs.
Theoretical implications
The formal model assumes each session is generated by a latent actor class Y \in H,A (human vs agent). Your deployment choices implicitly assert which sessions are valid for estimating human vs agent behavior and whether experimental conditions are stable. If you mix exploratory QA traffic with labeled experiments without recording that fact, you blur the empirical partitions D_H and D_A that the methodology needs for transition kernels and contamination studies. See the Introduction (research questions) and Methodology, Problem Formalization, in the thesis PDF.
2. Business fit framing
The problem PHANTOM addresses: Session-based pricing accumulates demand signals across a user's browsing history and raises quoted prices accordingly—the Cost of Information (COI) premium. LLM agents undercut this by separating reconnaissance (many isolated sessions, no signal accumulation) from execution (a clean session that quotes a floor price). The thesis proves that as the number of independent querying agents grows, the realizable price collapses to a minimum order statistic and COI approaches zero.
What PHANTOM gives you: A controlled platform to measure how much COI is at risk under real agent traffic, simulate that risk across contamination levels \alpha \in [0,1], and train pricing policies that remain robust. The pipeline runs from raw interaction logs through behavioral kernel estimation and a contamination generator to a DR-RL gym.
What you must supply:
- A product catalog path: defaults assume Supabase-backed product data (
NEXT_PUBLIC_SUPABASE_URL,NEXT_PUBLIC_SUPABASE_ANON_KEY). - A plan for interaction and price events reaching the ingestion path (backend → Kafka) or an adapter you maintain.
- Clear experiment goals: e.g. compare human vs agent KPIs under the same task, measure margin under varying contamination
\alpha.
Theoretical implications
Aggregate demand in the thesis is a mixture over human and agent types with contamination \alpha plus noise \epsilon_t; see the mixture demand discussion in Chapter 3 (Methodology). COI is defined as \mathbb{E}[P]-\underline{p}; the COI framework and theorem in the same chapter explain why saturated agent querying collapses extractable premium. Your business scenario determines which actions enter \hat{q} and how interpretable \alpha is for your traffic.
3. Environment and secrets
Bootstrap files (from repo root):
npm install
cp .env.example .env
cp .env.sweep.example .env.sweep
Core .env (platform + web + docker): See [.env.example](.env.example). You must also set the variables called out in [README.md](README.md) for a full stack: NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, AIRFLOW_FERNET_KEY, AIRFLOW_SECRET_KEY (and provider ports per your compose file).
Training / sweeps (.env.sweep): Used by make train, make benchmark, sweep agents. Typically WANDB_API_KEY, optional WANDB_ENTITY / WANDB_PROJECT, GITHUB_TOKEN for bootstrap flows, SWEEP_ID for W&B sweep workers. See [.env.sweep.example](.env.sweep.example).
Security: Never commit real .env or .env.sweep files. Rotate keys if they leak.
Theoretical implications
Splitting online platform credentials (ingestion, catalog, Kafka) from offline training credentials (W&B, cloud TPUs, GitHub tokens for workers) mirrors the hybrid Kappa–Lambda data loop in the thesis: streaming observation vs batch / long-running training jobs. That split is named in the Terminology appendix of the thesis PDF.
4. Bring-up (commands)
Aligned with [README.md](README.md):
npm install
cp .env.example .env
cp .env.sweep.example .env.sweep
# edit .env: Supabase, Airflow keys, etc.
make platform.up
make web.dev
Sanity checks:
| Endpoint | Role |
|---|---|
http://localhost:3000 |
Next.js storefront |
http://localhost:5000/health |
Backend ingest API |
http://localhost:5001/health |
Pricing provider |
http://localhost:8085 |
Airflow UI (default compose port) |
http://localhost:8084 or configured REDPANDA_CONSOLE_PORT |
Kafka console (see your .env) |
Optional tests: make test.backend (with venv/tooling as in Makefile); make test.e2e requires backend, web, and Airflow up per README.
Theoretical implications
A correctly wired stack logs trajectories \tau_s (sequences of events) and price exposure together. Chapter 3 defines events e_{s,k}=(a,i,t) and proxies \hat{q} from weighted actions—without joint logging of behavior and quotes, you cannot recover the objects the theory reasons about (Problem Formalization).
5. Service map
flowchart LR
U[Human / Agent Browser] --> W[Next.js Web App]
W -->|Price requests| P[Pricing Provider]
W -->|Interaction events| B[Backend Ingest API]
B --> K[Kafka]
K --> A[Airflow + Worker Jobs]
A --> R[Redis Model Registry]
P -->|Session/global prices| W
E[Research Engine + Experiments] --> A
E --> R
Ports (typical; confirm in docker-compose and .env): BACKEND_PORT (5000), PROVIDER_PORT (5001), KAFKA_PORT, REDIS_PORT, Airflow AIRFLOW_WEBSERVER_PORT (8085 default), Redpanda console.
Theoretical implications
The platform observes behavioral proxies and quoted prices, not the latent demand curve d(p\mid\theta). The distinction between \hat{q} and true demand is explicit in Chapter 3. Misattributing proxy noise to “true” elasticity breaks both estimation and any causal story about COI.
6. Tailoring to your business
Storefront mode: STORE_MODE=hotel or airline (see [web/src/lib/config.ts](web/src/lib/config.ts) and env). This switches catalog and UI, not the core ingestion pattern.
API base / environment: NEXT_PUBLIC_API_BASE, NEXT_PUBLIC_APP_ENV (validated in config.ts).
Paths for data and runs: Override with PHANTOM_DATA_DIR, PHANTOM_SIM_RUNS_DIR, PHANTOM_MODEL_REGISTRY_DIR, PHANTOM_COLLECTED_DATA_DIR, etc. ([lib/config.py](lib/config.py)).
Scope: A new vertical (custom product ontology, checkout rules, pricing rules) means new UI, events, and possibly new reward features in the engine. Budget engineering time; the repo is a research platform, not a turnkey SaaS skin for arbitrary catalogs without code changes.
Theoretical implications
Transition kernels \hat{\mathcal{T}}_H,\hat{\mathcal{T}}_A are estimated on a finite action / state space derived from your instrumentation. Changing catalog depth or event taxonomy changes the MDP state space; old kernel estimates are not portable. See the transition kernel discussion in Chapter 3.
7. Data collection and experiments
Flow: Browser → backend → Kafka → downstream consumers (Airflow DAGs, notebooks, ETL under experiments/). Ensure session identity, item identifiers, and action types are consistent enough to build trajectories.
Weak labels: The thesis discusses partitioning data into human vs agent subsets for MLE transition counts. In production you may only have heuristic labels—document bias explicitly.
Theoretical implications
Distinguishability (sub-question SQ1 in the Introduction) asks whether H vs A is identifiable from behavior alone. Your labeling and experimental design determine whether \Delta_H,\Delta_A and f(\tau) are meaningful or dominated by noise. Symbols appear in the Terminology appendix (\Delta_H,\Delta_A, f(\tau), contamination generator \mathcal{G}(\alpha)).
8. Transition kernels and agent scoring (theory → practice)
Theory: Sessions yield trajectories \tau_s. For each actor class y\inH,A, the thesis estimates a Markov transition kernel by counting transitions and normalizing (MLE):
\hat{P}(s' \mid s) = \frac{N(s,s')}{\sum_k N(s,k)}
Human and agent prototypes \hat{\mathcal{T}}_H,\hat{\mathcal{T}}_A support comparing an empirical kernel from a partial trajectory to prototypes (e.g. KL-style divergences \Delta_H,\Delta_A) and mapping to a weak agent probability f(\tau). See Chapter 3 and the Terminology appendix.
Code: [engine/lib/coi.py](engine/lib/coi.py) (compute_agent_probability: empirical transition counts vs human/agent reference dicts, KL-style terms, mapped via [lib/agent_probability.py](lib/agent_probability.py)).
Optional narrative: [blog/02-behavioral-fingerprinting.md](blog/02-behavioral-fingerprinting.md) walks a concrete study design (not required for operators).
Theoretical implications
If reference kernels are fit on stale or mislabeled partitions, \Delta_H-\Delta_A is not interpretable as distinguishability. Ground claims in SQ1 (Introduction) and the kernel subsection of Chapter 3.
9. Contamination generator \mathcal{G}(\alpha)
Theory: Given clean trajectories, \mathcal{G}(\alpha) injects synthetic agent trajectories until the effective mixture reaches contamination \alpha\in[0,1], defining training scenarios for robust policies (Chapter 3). Catalog-scale block expansion of kernels is discussed there with validation caveats—treat large product spaces as research-grade until your team signs off.
Code: [engine/engine.py](engine/engine.py) — MarketEngine mixes human/agent demand, uses get_adjusted_transitions / sample_behavior_from_transitions, and alpha when combining actor types and building demand proxies (estimate_demand). This is the simulator path, not a drop-in replacement for your production database.
Theoretical implications
\alpha in mixture Q(p) is agentic demand contribution in the formal model, not necessarily “bot share of page views” unless your instrumentation equates them. Mismeasuring \alpha biases robust objectives tied to a fixed contamination level.
10. Training and evaluation — local workflow
Environment: Python venv via Nx (make install / nx run research:install). Training commands load .env.sweep.
make train LOCAL_TRAIN_ARGS='--algo ppo --total-timesteps 50000'
make benchmark LOCAL_BENCHMARK_ARGS='--tiers static,surge,linear,qtable,ppo --alpha-values 0.0,0.3 --episodes 3 --no-wandb'
make benchmark.simple
Entrypoints: [engine/train.py](engine/train.py), [engine/benchmark.py](engine/benchmark.py), [engine/spec.py](engine/spec.py) (Nx wraps these—see project.json / research targets).
Artifacts: [lib/config.py](lib/config.py) — PHANTOM_SIM_RUNS_DIR (default sim/rl/runs), PHANTOM_MODEL_REGISTRY_DIR, etc.
TensorBoard (optional): [docker-compose.yml](docker-compose.yml) includes tensorboard-rl on host port 6007 (./sim/rl/runs) and tensorboard-ml on 6006 (./experiments/ml/runs).
Theoretical implications
Local runs instantiate the offline defense gym: policies trained on simulator-induced distributions approximate the DR-RL narrative in Chapter 3, but hyperparameters (\lambda on COI leakage, \eta on UX, robust radius) change the effective ambiguity set. Cross-check engine/ against the thesis before claiming figure-for-figure replication.
11. Training and evaluation — remote / scaled deployment
For research at scale (cloud quota and secrets required):
| Mechanism | Role |
|---|---|
[submit_ray_job.sh](submit_ray_job.sh) |
Ray jobs with .env injected; `RAY_MODE=single |
make tpu.ray.bootstrap / tpu.ray.* |
TPU Ray bootstrap (TPU_CONF, e.g. tpu_orchestration/configs/v4_spot_us.conf). |
make train.agent / make benchmark.agent |
W&B sweeps: SWEEP_ID in .env.sweep. |
make train.bootstrap |
Worker bootstrap: REPO_URL, SWEEP_ID, GITHUB_TOKEN. |
make docker.train.publish |
Trainer image (TRAIN_IMAGE_REF in Makefile). |
See submit_ray_job.sh for env vars (WANDB_*, PHANTOM_* TPU toggles).
Theoretical implications
Distributed training does not change the definitions of the Stackelberg game or Wasserstein ambiguity; it changes compute and variance of empirical estimates. Align random seeds and data protocol across nodes or split results explicitly—otherwise you mix distributions in a way a single empirical law \hat{P}_N in the thesis does not describe.
12. Evaluation, artifacts, and audit trail
Benchmarks: make benchmark* sweeps tiers and \alpha; CLI includes robustness knobs (see default BENCHMARK_ARGS in submit_ray_job.sh: --robust-radius, --lambda-coi, --eta-ux, etc.).
Audit trail: Store git SHA, CLI argv, non-secret .env.sweep keys, and W&B run IDs with published tables. For scientific claims, cite Chapters 4–5 (Results, Discussion) in the thesis PDF.
Theoretical implications
Evaluation quality equals simulator fidelity plus contamination modeling. Separate theorem statements (assumption-based) from empirical curves (engine-dependent).
13. Operational suggestions
- Staging: Non-production namespaces; separate Kafka topics and Supabase projects where possible.
- Rate limits / abuse: Protect ingest endpoints; respect participant privacy.
- Human vs agent sessions: Comparable cohorts; record experimental condition in metadata.
- Contracts:
tests/e2e/encodes minimal flows—use when APIs change.
Theoretical implications
Non-stationary noise \epsilon_t and drifting \alpha confound benchmark interpretation. Chapter 3 discusses mixture identification: isolate treatments when possible and document confounders when not.
14. Roadmap and gaps
In repo: Local dockerized stack, demo verticals, engine benchmarks, documented env and paths.
Usually custom: Production catalog without Supabase, identity/fraud layers, legal review of logging, Kafka/Airflow SLAs, hardening the pricing provider for real money.
Thesis vs code: The PDF is the spec; not every robustness term or large-catalog kernel construction is production-verified—see caveats in Chapter 3.
Theoretical implications
Theorems in the thesis can be stronger than what observational firm logs support. The COI result assumes a clean experimental reading of the pricing policy; live market data may only support weaker claims.
15. Theory and thesis cross-references (quick index)
Use the PDF table of contents with these anchors:
| Topic | Thesis location |
|---|---|
| Research questions (margin, distinguishability, contamination, mitigation) | Introduction |
Sessions, events, \hat{q}, mixture Q(p), \alpha |
Chapter 3 — Problem Formalization, mixture demand |
| COI definition and erosion theorem | Chapter 3 — COI framework |
Transition kernels, MLE, \mathcal{G}(\alpha) |
Chapter 3 |
| DR-RL, ambiguity sets, Stackelberg | Chapter 3 |
Symbol glossary (COI leakage, f(\tau), UX, surrogates) |
Appendix — Terminology |
| Empirical results and limitations | Chapters 4–5 |
16. Quick file index (code)
| File | Role |
|---|---|
[engine/lib/coi.py](engine/lib/coi.py) |
KL-style trajectory comparison; agent probability. |
[engine/engine.py](engine/engine.py) |
MarketEngine, mixture, demand proxy path. |
[lib/agent_probability.py](lib/agent_probability.py) |
Divergence → probability score. |
[lib/config.py](lib/config.py) |
Paths and ports for artifacts. |
[engine/train.py](engine/train.py), [engine/benchmark.py](engine/benchmark.py) |
CLI entrypoints. |
[tpu_orchestration/](tpu_orchestration/) |
TPU configs and helpers. |
Many offline benchmarks run without a storefront once the research Python environment is installed; connecting production trajectories to kernel estimation still requires aligned instrumentation.