feat: baseline setup for RL modeling

experiments/procesing/contaminator.py

@@ -1,7 +1,13 @@
-import pandas as pd
-import random
+from __future__ import annotations
+
 import os
+import random
 from pathlib import Path
+from types import SimpleNamespace
+
+import pandas as pd
+
+from lib.separability import estimate_alpha, load_artifacts, score_session
 
 # use relative import when in package context, fallback for standalone
 try:
@@ -15,6 +21,11 @@ except ImportError:
 PROJECT_ROOT = Path(__file__).parent.parent.parent
 AGENT_DATA_DIR = Path(os.getenv('PHANTOM_AGENT_DATA_DIR', PROJECT_ROOT / "experiments" / "agents" / "collected_data"))
 
+try:
+    SEPARABILITY_ARTIFACTS = load_artifacts()
+except FileNotFoundError:
+    SEPARABILITY_ARTIFACTS = None
+
 
 def remap_schema(df: pd.DataFrame, mapping: dict, on: str = "event_type") -> pd.DataFrame:
     """remap column values according to mapping dict, preserving unmapped values"""
@@ -23,6 +34,24 @@ def remap_schema(df: pd.DataFrame, mapping: dict, on: str = "event_type") -> pd.
     return df
 
 
+def _states_to_events(states: list[str]) -> list[SimpleNamespace]:
+    events: list[SimpleNamespace] = []
+    for idx, state in enumerate(states):
+        parts = state.split("|") if isinstance(state, str) else ["page", "product", str(state)]
+        page = f"/{parts[0]}" if parts else "/"
+        product = parts[1] if len(parts) > 1 else "unknown"
+        event_name = parts[2] if len(parts) > 2 else parts[-1]
+        events.append(
+            SimpleNamespace(
+                eventName=event_name,
+                page=page,
+                productId=product,
+                ts=float(idx),
+            )
+        )
+    return events
+
+
 def contaminate_dataset(df: pd.DataFrame, on: str = "event_type",
                         contamination_rate: float = 0.1,
                         agent_data_dir: Path = None) -> pd.DataFrame:
@@ -48,6 +77,7 @@ def contaminate_dataset(df: pd.DataFrame, on: str = "event_type",
 
     # generate synthetic trajectories
     new_rows = []
+    alpha_estimates = []
     for start_event in start_events:
         # sample trajectory from agent model, using a state that contains the event type
         mdp_states = model.mdp.get('states', []) if model.mdp else []
@@ -56,11 +86,28 @@ def contaminate_dataset(df: pd.DataFrame, on: str = "event_type",
             continue  # skip if no matching start state
         start_state = random.choice(matching_starts)
         trajectory = model.sample_traj(start_state, max_len=20)
+        score_payload: list[SimpleNamespace] = []
+        score: dict[str, float] = {}
+        if SEPARABILITY_ARTIFACTS:
+            score_payload = _states_to_events(trajectory)
+            score = score_session(score_payload, SEPARABILITY_ARTIFACTS)
+            alpha_estimates.append(
+                estimate_alpha(score["prob_agent"], score["delta_h"], score["delta_a"], temperature=2.0)
+            )
+
         for state in trajectory:
-            parts = state.split('|')  # page|productId|eventName format
-            new_rows.append({on: parts[-1] if parts else start_event, 'source': 'synthetic_agent'})
+            parts = state.split('|') if isinstance(state, str) else [start_event]
+            new_rows.append({
+                on: parts[-1] if parts else start_event,
+                'source': 'synthetic_agent',
+                'prob_agent': score.get('prob_agent') if SEPARABILITY_ARTIFACTS and score_payload else None,
+                'delta_h': score.get('delta_h') if SEPARABILITY_ARTIFACTS and score_payload else None,
+                'delta_a': score.get('delta_a') if SEPARABILITY_ARTIFACTS and score_payload else None,
+            })
 
     if new_rows:
         contaminate_df = pd.DataFrame(new_rows)
         df = pd.concat([df, contaminate_df], ignore_index=True)
+        if alpha_estimates:
+            df['estimated_alpha'] = sum(alpha_estimates) / len(alpha_estimates)
     return df