| """Re-simulate recorded episodes from the `steps` table and check that the environment reproduces them exactly. |
| |
| uv run python scripts/replay.py # the best episode of the evaluation |
| uv run python scripts/replay.py --episodes gpt-6.1-sol@medium/bursill_bunny claude-haiku-5.5@max/trewey_swan |
| uv run python scripts/replay.py --all # all 360 episodes (about two days on one core) |
| |
| Each episode is replayed through a fresh environment on its certified target. The script checks that the first |
| observation and every image a call returned are pixel-identical to the recorded ones, and that every call is |
| accepted or refused as recorded, returns the recorded text, leaves the pose that hashes to `pose_hash` and has the |
| recorded hidden metrics. |
| """ |
| import os |
|
|
| os.environ.setdefault("OPENBLAS_NUM_THREADS", "1") |
|
|
| import argparse |
| import io |
| import json |
| from pathlib import Path |
|
|
| import pyarrow.compute as pc |
| import pyarrow.parquet as pq |
| from PIL import Image, ImageChops |
|
|
| from shadowbench.factory import create |
| from shadowbench.native_environment import state_hash |
|
|
| DATA = Path(__file__).resolve().parents[1] / "data" |
| METRICS = ("hidden_score", "score_before_peeks", "iou", "boundary_f1", "hole_preservation", "ease") |
|
|
|
|
| def same_image(a: Image.Image, recorded: dict) -> bool: |
| with Image.open(io.BytesIO(recorded["bytes"])) as b: |
| return a.size == b.size and ImageChops.difference(a.convert("RGB"), b.convert("RGB")).getbbox() is None |
|
|
|
|
| def replay(episode_id: str) -> int: |
| model, rest = episode_id.split("@", 1) |
| effort, task_id = rest.split("/", 1) |
| table = pq.read_table(DATA / "steps" / f"{effort}_effort" / f"{model}.parquet") |
| rows = sorted(table.filter(pc.equal(table["episode_id"], episode_id)).to_pylist(), key=lambda s: s["step"]) |
| if not rows: |
| raise SystemExit(f"no steps for {episode_id}") |
| episode = next(x for x in pq.read_table(DATA / "episodes" / f"{effort}_effort.parquet").to_pylist() |
| if x["episode_id"] == episode_id) |
| env = create(task_id, revision=episode["target_revision"]) |
| if env.version != episode["env_fingerprint"]: |
| raise SystemExit(f"environment fingerprint {env.version} differs from the recorded {episode['env_fingerprint']}; " |
| "install the pinned versions with `uv sync`") |
| problems = [] |
| if state_hash(env.pose) != rows[0]["pose_hash"] or not same_image(env.observe(), rows[0]["image"]): |
| problems.append("first observation") |
| for s in rows[1:]: |
| out = env.call(s["tool"], json.loads(s["arguments"])) |
| m = env.metrics() |
| got = {"hidden_score": m["score"], "score_before_peeks": m["score_before_peeks"], "iou": m["iou"], |
| "boundary_f1": m["boundary_f1"], "hole_preservation": m["hole_preservation"], "ease": m["ease"]} |
| if env.steps[-1]["accepted"] != s["accepted"]: |
| problems.append(f"step {s['step']}: accepted {env.steps[-1]['accepted']}, recorded {s['accepted']}") |
| if state_hash(env.pose) != s["pose_hash"]: |
| problems.append(f"step {s['step']}: pose differs") |
| if out.text != s["response"]: |
| problems.append(f"step {s['step']}: response differs") |
| if any(got[k] != s[k] for k in METRICS): |
| problems.append(f"step {s['step']}: metrics differ " + str({k: (got[k], s[k]) for k in METRICS if got[k] != s[k]})) |
| if (out.image is None) != (s["image"] is None) or (out.image is not None and not same_image(out.image, s["image"])): |
| problems.append(f"step {s['step']}: image differs") |
| if not env.done or env.final["score"] != episode["score"]: |
| problems.append(f"final score {env.final and env.final['score']} != recorded {episode['score']}") |
| n = len(rows) - 1 |
| print(f"{episode_id}: {n} call{'' if n == 1 else 's'} replayed, " |
| + ("all states, decisions, texts, images and scores identical" if not problems else f"{len(problems)} differences"), |
| flush=True) |
| for p in problems[:20]: |
| print(" ", p) |
| return len(problems) |
|
|
|
|
| ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) |
| ap.add_argument("--episodes", nargs="+", default=["claude-opus-5.5@max/hilliar_giraffe"]) |
| ap.add_argument("--all", action="store_true") |
| a = ap.parse_args() |
| ids = a.episodes |
| if a.all: |
| ids = [x for e in ("medium", "high", "max") for x in |
| pq.read_table(DATA / "episodes" / f"{e}_effort.parquet", columns=["episode_id"]).column(0).to_pylist()] |
| bad = sum(replay(i) for i in ids) |
| raise SystemExit(1 if bad else 0) |
|
|
