1 files changed, 127 insertions, 0 deletions
diff --git a/experiments/analyze_babyai_shared_b1.py b/experiments/analyze_babyai_shared_b1.py
new file mode 100644
index 0000000..ebcdb17
--- /dev/null
+++ b/experiments/analyze_babyai_shared_b1.py
@@ -0,0 +1,127 @@
+#!/usr/bin/env python3
+"""Summarize the frozen BabyAI GoToObj shared-feedback endpoint."""
+
+import argparse
+import json
+from pathlib import Path
+import statistics
+
+import gymnasium as gym
+import minigrid  # noqa: F401 - registers BabyAI environments
+
+
+ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_RESULTS = ROOT / "results" / "babyai_shared" / "b1"
+DEFAULT_OUT = ROOT / "results" / "babyai_shared" / "b1_analysis.json"
+SEEDS = (4101, 4102, 4103)
+CONDITIONS = ("bp", "clean_kp", "raw_shared", "sdil")
+
+
+def summarize(values):
+    return {
+        "values": values,
+        "mean": statistics.mean(values),
+        "sample_std": statistics.stdev(values),
+    }
+
+
+def object_counts(env_id, seeds):
+    counts = []
+    for seed in seeds:
+        env = gym.make(env_id)
+        try:
+            env.reset(seed=seed)
+            counts.append(sum(
+                cell is not None and cell.type in {"ball", "box", "key"}
+                for cell in env.unwrapped.grid.grid))
+        finally:
+            env.close()
+    return counts
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--results", type=Path, default=DEFAULT_RESULTS)
+    parser.add_argument("--out", type=Path, default=DEFAULT_OUT)
+    args = parser.parse_args()
+    records = {}
+    for seed in SEEDS:
+        for condition in CONDITIONS:
+            path = args.results / f"seed{seed}_{condition}.json"
+            with open(path, encoding="utf-8") as handle:
+                records[(seed, condition)] = json.load(handle)
+
+    condition_summary = {}
+    for condition in CONDITIONS:
+        condition_records = [records[(seed, condition)] for seed in SEEDS]
+        condition_summary[condition] = {
+            "rollout_success_percent": summarize([
+                100.0 * row["rollout"]["success"]
+                for row in condition_records]),
+            "expert_action_accuracy_percent": summarize([
+                100.0 * row["validation"]["accuracy"]
+                for row in condition_records]),
+            "mission_lesion_rollout_success_percent": summarize([
+                100.0 * row["mission_lesion_rollout"]["success"]
+                for row in condition_records]),
+            "training_wall_seconds": summarize([
+                row["training"]["training_wall_seconds"]
+                for row in condition_records]),
+            "all_finite": all(row["finite"] for row in condition_records),
+        }
+    rollout_gaps = [100.0 * (
+        records[(seed, "sdil")]["rollout"]["success"]
+        - records[(seed, "raw_shared")]["rollout"]["success"])
+                    for seed in SEEDS]
+    action_gaps = [100.0 * (
+        records[(seed, "sdil")]["validation"]["accuracy"]
+        - records[(seed, "raw_shared")]["validation"]["accuracy"])
+                   for seed in SEEDS]
+    counts = object_counts("BabyAI-GoToObjS6-v1", range(256))
+    task_required = min(counts) > 1
+    checks = {
+        "all_runs_finite": all(
+            row["finite"] for row in records.values()),
+        "bp_and_clean_kp_mean_rollout_at_least_80": (
+            condition_summary["bp"]["rollout_success_percent"]["mean"] >= 80
+            and condition_summary["clean_kp"]
+            ["rollout_success_percent"]["mean"] >= 80),
+        "sdil_rollout_above_raw_every_seed": all(
+            value > 0 for value in rollout_gaps),
+        "mission_structurally_required": task_required,
+    }
+    report = {
+        "stage": "babyai_shared_b1_analysis",
+        "gate": "pass" if all(checks.values()) else "fail",
+        "checks": checks,
+        "conditions": condition_summary,
+        "paired_sdil_minus_raw": {
+            "rollout_success_points": summarize(rollout_gaps),
+            "expert_action_accuracy_points": summarize(action_gaps),
+        },
+        "task_structure_audit": {
+            "env_id": "BabyAI-GoToObjS6-v1",
+            "audited_episode_seeds": 256,
+            "minimum_task_object_count": min(counts),
+            "maximum_task_object_count": max(counts),
+            "mission_structurally_required_to_identify_the_only_object": (
+                task_required),
+            "interpretation": (
+                "The sole object is the target, so the mission does not select "
+                "among alternatives. Mission lesion is an input ablation, not "
+                "evidence that language is required by the task."),
+        },
+        "test_split_generated_or_read": False,
+        "decision": (
+            "Retain as an implementation diagnostic; do not use as the paper's "
+            "task-required shared-feedback result. Move to PickupLoc."),
+    }
+    args.out.parent.mkdir(parents=True, exist_ok=True)
+    with open(args.out, "w", encoding="utf-8") as handle:
+        json.dump(report, handle, indent=2, sort_keys=True)
+        handle.write("\n")
+    print(json.dumps(report, indent=2, sort_keys=True))
+
+
+if __name__ == "__main__":
+    main()