EXPERIMENTS 1–6: ORIGINAL RESULTS AND FRESH REPLICATION
Updated: 2026-11-03T18:12:25.718841+00:00

Archived = authors' saved result. Released rerun = original policy weights evaluated again. Fresh = newly generated Codex answer/policy.

Experiment 1: game identification
Archived 4/4; fresh 4/4 by the released category rule. Both accept Frogger as Freeway, so this is not four exact game titles.

Experiment 2: diffusion shape identification
Archived 3/3 shape categories; fresh 3/3. Both incorrectly described hollow outlines in all three cases, rather than filled shapes.

Experiment 3: language networks
Measure | Archived | Fresh
Corpus identification | 6/6 | 3/6
Unshuffled token order | 6/6 | 6/6
Exact copy-head set | 6/6 | 5/6
Fresh English: 2/2; Python: 1/2; romaji: 0/2. Copy-head criterion: released induction metric >0.2.

Experiment 4: car repair
Policy | Test successes / 1024
Authors archived repair | 1024
Released repair rerun | 1024
Fresh repair | 0
The released repair takes 73 steps on average. Every fresh repair episode ends in a red collision; average length 33.15625 steps. The rounded unedited network fails by timeout.

Experiment 5: MinAtar
Game | Archived mean | Released rerun mean +/- SE | Fresh mean +/- SE
breakout | 517.411 | 496.270 +/- 7.817 | 380.585 +/- 6.176
seaquest | 53.110 | 51.533 +/- 1.348 | 35.245 +/- 0.842
asterix | 53.178 | 52.202 +/- 0.992 | 64.378 +/- 0.942
1000 episodes per policy; 6000 total. Fresh Asterix improves; fresh Breakout and Seaquest decline.

Experiment 6: Go
Mean win rate against released same-size PPO seeds 0 and 2, 4000 games per policy:
Passes | Archived against same opponents | Released rerun | Fresh | Fresh minus rerun (percentage points)
0 | 39.400% | 40.275% | 19.500% | -20.775
10 | not reported | 49.825% | 69.375% | +19.550
25 | not reported | 47.950% | 69.750% | +21.800
50 | 76.350% | 77.225% | 85.125% | +7.900

Original headline Go means (against DIFFERENT, unavailable smaller opponents):
0 passes: 54.525%
10 passes: 66.400%
25 passes: 69.125%
50 passes: 82.400%
These headline means cannot be directly compared with the reconstructed same-size-opponent evaluation.
48,000 held-out Go games total across four budgets, two policies and three opponents (including random).

Limits
Original prompts and released inputs retained, except disclosed Experiment 6 opponent-description changes. Historical CLI/backend and some original evaluator dependencies are unavailable.
Car and MinAtar use reconstructed wrappers and new RNG streams.
Go uses released same-size PPO seeds 0 and 2. The original headline results use unavailable smaller PPO opponents; those headline rates are not directly comparable.
Go archived same-size comparisons are available for budgets 0 and 50. Their original RNG and wrapper differ from our rerun.
Single fresh writer trial per task/budget; game-level error bars do not measure variation between new Codex trials.
Original Experiment 5 PPO comparison checkpoints are unavailable. No full PPO baseline retraining was launched; it is cancelled.

Recorded episode statistics and artifact hashes passed independent checks. This is not a second full simulator replay.
Machine-readable comparison and source hashes: experiments-1-through-6-comparison.json
