| """Reproduce the leaderboard and the report's headline numbers from the episode and task tables. |
| |
| uv run python scripts/analyze.py |
| |
| Scores are rounded to two decimals per episode, as in the report, and the 95% intervals are bootstrap intervals over |
| the 20 plates (10,000 resamples, seed 20261105, drawn in the report's order), so every number printed here matches |
| the report and the `leaderboard` config exactly. The script exits with an error if one does not. |
| """ |
| import json |
| from pathlib import Path |
|
|
| import numpy as np |
| import pyarrow.parquet as pq |
|
|
| DATA = Path(__file__).resolve().parents[1] / "data" |
| MODELS = ["claude-opus-5.5", "claude-sonnet-5.5", "gpt-6.1-sol", "gpt-6.1-terra", "claude-haiku-5.5", "gpt-6-luna"] |
| EFFORTS = ["medium", "high", "max"] |
| rng = np.random.default_rng(20261105) |
|
|
|
|
| def r(x, d=2): |
| return float(round(float(x), d)) |
|
|
|
|
| def boot(v): |
| m = v[rng.integers(0, len(v), size=(10000, len(v)))].mean(axis=1) |
| return r(np.percentile(m, 2.5)), r(np.percentile(m, 97.5)) |
|
|
|
|
| def read(path): |
| return pq.read_table(path).to_pylist() |
|
|
|
|
| episodes = [x for e in EFFORTS for x in read(DATA / "episodes" / f"{e}_effort.parquet")] |
| tasks = {t["task_id"]: t for t in read(DATA / "tasks" / "tasks.parquet")} |
| board = {(x["model"], x["effort"]): x for x in read(DATA / "leaderboard" / "leaderboard.parquet") if x["kind"] == "model"} |
| refs = {x["model"]: x for x in read(DATA / "leaderboard" / "leaderboard.parquet") if x["kind"] == "reference"} |
| plates = sorted(tasks) |
| name = {x["model"]: x["model_name"] for x in episodes} |
|
|
|
|
| def pick(m=None, e=None): |
| return sorted((x for x in episodes if (m is None or x["model"] == m) and (e is None or x["effort"] == e)), |
| key=lambda x: x["task_id"]) |
|
|
|
|
| rows, failures = {}, [] |
| for m in MODELS: |
| for e in EFFORTS: |
| xs = pick(m, e) |
| s = np.array([r(x["score"]) for x in xs]) |
| lo, hi = boot(s) |
| rows[(m, e)] = dict(score=r(s.mean()), ci_low=lo, ci_high=hi, |
| cost_usd_per_episode=r(np.mean([x["cost_usd"] for x in xs]), 4), |
| minutes_per_episode=r(np.mean([x["seconds"] for x in xs]) / 60, 1), |
| plates_above_blob=sum(r(x["score_before_peeks"]) > r(tasks[x["task_id"]]["blob_score"]) for x in xs)) |
| for m in MODELS: |
| a = np.array([r(x["score"]) for x in pick(m, "medium")]) |
| b = np.array([r(x["score"]) for x in pick(m, "max")]) |
| lo, hi = boot(b - a) |
| rows[(m, "max")].update(gain_vs_medium=r((b - a).mean()), gain_ci_low=lo, gain_ci_high=hi, |
| plates_improved_vs_medium=int((b > a).sum())) |
| search = np.array([r(tasks[t]["search_score"]) for t in plates]) |
| slo, shi = boot(search) |
|
|
| for key, row in rows.items(): |
| for k, v in row.items(): |
| if board[key][k] != v: |
| failures.append(f"{key} {k}: {v} != {board[key][k]}") |
| if (r(search.mean()), slo, shi) != (refs["search"]["score"], refs["search"]["ci_low"], refs["search"]["ci_high"]): |
| failures.append("search reference differs") |
| for ref, col in (("blob", "blob_score"), ("rest", "rest_score")): |
| if r(np.mean([r(tasks[t][col]) for t in plates])) != refs[ref]["score"]: |
| failures.append(f"{ref} reference differs") |
|
|
| usd = lambda v: f"${v:.3f}" if v < 0.1 else f"${v:.2f}" |
| print("| Model | Medium | High | Max | Cost per episode (medium · high · max) |") |
| print("| --- | ---: | ---: | ---: | ---: |") |
| for m in sorted(MODELS, key=lambda m: -rows[(m, "max")]["score"]): |
| sc = [f"{rows[(m, e)]['score']:.1f} [{rows[(m, e)]['ci_low']:.1f}, {rows[(m, e)]['ci_high']:.1f}]" for e in EFFORTS] |
| cost = " · ".join(usd(rows[(m, e)]["cost_usd_per_episode"]) for e in EFFORTS) |
| print(f"| {name[m]} | {' | '.join(sc)} | {cost} |") |
| print() |
| print("| Model | Max minus medium [95% CI] | Plates better at max |") |
| print("| --- | ---: | ---: |") |
| for m in sorted(MODELS, key=lambda m: -rows[(m, "max")]["gain_vs_medium"]): |
| g = rows[(m, "max")] |
| print(f"| {name[m]} | {g['gain_vs_medium']:+.1f} [{g['gain_ci_low']:+.1f}, {g['gain_ci_high']:+.1f}] | {g['plates_improved_vs_medium']} of 20 |") |
| print() |
| print(f"References: blob {refs['blob']['score']:.1f}, blind search {r(search.mean()):.1f} [{slo:.1f}, {shi:.1f}], " |
| f"hands at rest {refs['rest']['score']:.1f}") |
| two = [x for x in episodes if x["both_hands_cast_shadow"]] |
| moves = np.array([[x["moves_better"], x["moves_worse"], x["moves_unchanged"]] for x in episodes]).sum(axis=0) |
| print(f"Episodes: {len(episodes)}, solved {sum(x['solved'] for x in episodes)}, total cost {usd(sum(x['cost_usd'] for x in episodes))}") |
| print(f"Final poses with both hands casting a shadow: {len(two)}, hands together: " |
| f"{sum(1 for x in two if x['hand_gap_cm'] < 2 and x['hand_depth_gap_cm'] < 5)}, " |
| f"median gap {np.median([x['hand_gap_cm'] for x in two]):.1f} cm, median depth gap {np.median([x['hand_depth_gap_cm'] for x in two]):.1f} cm") |
| print(f"Accepted hand moves: {moves.sum()}, better {moves[0]}, worse {moves[1]}, unchanged {moves[2]}") |
| if failures: |
| raise SystemExit("MISMATCH\n" + "\n".join(failures)) |
| print("\nAll leaderboard numbers reproduced exactly.") |
|
|
