File size: 2,069 Bytes
4db68fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
"""Claim 2: empirical predictability ceiling (paper Section 2.1, Figure 3).

Paper: across the 52 trajectories, 68% of all ground-truth (cell, property)
pairs are predicted by at least one frontier-LLM oracle (union over Opus 4.6,
Opus 4.7, GPT-5.4, GPT-5.5, 2 generations per step); per-trajectory median
66.3%, 44/52 trajectories exceed 50%. Figure 3 reports mean 66.0.

The repo ships the oracle output per trajectory as
data/raw/<id>/predictable_state.json with fields predictable_count,
final_state_size, coverage_pct. We aggregate those.
"""
import glob
import json
import statistics

files = sorted(glob.glob("NAPE/data/raw/*/predictable_state.json"))
rows = []
for f in files:
    d = json.load(open(f))
    rows.append((d["trajectory_name"], d["predictable_count"],
                 d["final_state_size"], d["coverage_pct"]))

n = len(rows)
tot_pred = sum(r[1] for r in rows)
tot_state = sum(r[2] for r in rows)
covs = [r[3] for r in rows]
overall = 100 * tot_pred / tot_state
over50 = sum(c > 50 for c in covs)

print(f"trajectories with oracle file : {n}")
print(f"total predictable properties  : {tot_pred}")
print(f"total final-state properties  : {tot_state}")
print(f"overall coverage (pooled)     : {overall:.1f}%   (paper: 68%)")
print(f"per-trajectory mean           : {statistics.mean(covs):.1f}%   (paper Fig.3: 66.0)")
print(f"per-trajectory median         : {statistics.median(covs):.1f}%   (paper: 66.3%)")
print(f"trajectories > 50% coverage   : {over50}/{n}  (paper: 44/52)")

json.dump({
    "n": n, "overall_pct": round(overall, 2),
    "mean_pct": round(statistics.mean(covs), 2),
    "median_pct": round(statistics.median(covs), 2),
    "over_50": over50,
    "per_trajectory": [{"name": r[0], "coverage_pct": r[3]} for r in rows],
}, open("repro_outputs/claim2_predictability.json", "w"), indent=2)
print("\nWrote repro_outputs/claim2_predictability.json")

ok = abs(overall - 68) < 1 and abs(statistics.median(covs) - 66.3) < 0.2 and over50 == 44
print("CLAIM 2 (from shipped oracle artifacts):", "REPRODUCED" if ok else "CHECK NUMBERS")