game stringclasses 4
values | seed int64 0 2 | steps int64 20M 20M | trained_mean float64 1.13 32.6 | trained_sd float64 1.94 9.93 | random_mean float64 -1.89 6.43 | random_sd float64 0.31 2.9 | diff float64 2.97 26.2 | welch_t float64 18.6 265 | passes_gate bool 1
class | wall_seconds float64 758 809 | config dict |
|---|---|---|---|---|---|---|---|---|---|---|---|
breakout | 0 | 19,988,480 | 3.462 | 2.464256 | 0.188 | 0.420305 | 3.274 | 29.256092 | true | 785.933058 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
breakout | 1 | 19,988,480 | 3.49 | 2.546743 | 0.198 | 0.450329 | 3.292 | 28.434092 | true | 781.535218 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
breakout | 2 | 19,988,480 | 3.308 | 2.506219 | 0.174 | 0.404628 | 3.134 | 27.576714 | true | 757.923325 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
pong | 0 | 19,988,480 | 1.598 | 2.423303 | -1.886 | 2.443973 | 3.484 | 22.612759 | true | 775.686345 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
pong | 1 | 19,988,480 | 1.216 | 2.514228 | -1.828 | 2.464633 | 3.044 | 19.313382 | true | 806.239366 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
pong | 2 | 19,988,480 | 1.132 | 2.547661 | -1.838 | 2.498751 | 2.97 | 18.59168 | true | 805.839616 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
pacman | 0 | 19,988,480 | 30.812 | 9.864312 | 6.16 | 2.712637 | 24.652 | 53.827688 | true | 805.988885 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
pacman | 1 | 19,988,480 | 31.576 | 9.925534 | 6.432 | 2.79381 | 25.144 | 54.472071 | true | 809.003253 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
pacman | 2 | 19,988,480 | 32.57 | 9.897934 | 6.38 | 2.897516 | 26.19 | 56.726634 | true | 790.505648 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
freeway | 0 | 19,988,480 | 23.466 | 2.031956 | 0.104 | 0.311743 | 23.362 | 253.859902 | true | 776.095872 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
freeway | 1 | 19,988,480 | 23.284 | 2.149266 | 0.094 | 0.305228 | 23.19 | 238.629962 | true | 765.055719 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
freeway | 2 | 19,988,480 | 23.458 | 1.939133 | 0.104 | 0.318094 | 23.354 | 265.483802 | true | 765.139078 | {
"total_steps": 20000000,
"n_envs": 128,
"n_steps": 128,
"epochs": 4,
"minibatches": 4,
"lr": 0.00025,
"gamma": 0.99,
"gae_lambda": 0.95,
"clip": 0.2,
"ent_coef": 0.01,
"vf_coef": 0.5,
"max_grad_norm": 0.5,
"eval_episodes": 500,
"gate_min_diff": 1,
"gate_min_t": 3.3
} |
Weight-Reading Probe
Can Claude tell what a small neural network was trained to do, just by reading its raw weights?
This is a small, blind, pre-registered partial replication of Experiment 1 from "Sparks of Weight-Space Programming" (Takehi et al., MIT CSAIL, Nov 2026). That study tested GPT-6 Astra. This one tests Claude.
- Designed and run by: Claude (Anthropic).
- Funded, and answer key held, by: @SoulFireMage.
Headline (see RESULTS.md)
| Hypothesis | Result |
|---|---|
| H1: identify the game from the weights | 12/12 forced choice on both attempts (chance 25%) |
| H2: evidence holds up | 36/36 quoted values and 35/38 structural claims verified |
| H0: says "cannot tell" on untrained or shuffled nets | 4/4 |
| H3: beats a dumb baseline | Fails. k-means and nearest-neighbour on first-layer weight mass also score 12/12 |
The honest summary: Claude identified every network and grounded its answers in real structure, but these networks turned out to be trivially easy to tell apart. Each game leaves a fingerprint in the weights: the set of cells its agent ever occupies. That means this test can't separate understanding from fingerprint-matching. A follow-up with matched-fingerprint games is proposed in RESULTS.md.
Files
- PREREGISTRATION.md: frozen before any training.
- DEVIATIONS.md: D1–D4, every one logged before unblinding.
prompts/: the reader prompts, frozen before reading.blinded/: the 16 networks as the readers saw them, plusKEY_HASH.txt.reads/: all 32 predictions, committed before unblinding.baselines/: k-means, committed before unblinding.- RESULTS.md: verdicts, evidence checks and interpretation.
- blind.py: the key-holder's blinding script.
Spend log (cap £10)
| Job | Hardware | Cost |
|---|---|---|
| Smoke test | cpu-upgrade | < $0.01 |
| Training (14.6 min) | cpu-xl ($1.00/h) | ≈ $0.24 |
| Leak check, k-means, verification (5 short jobs) | cpu-basic ($0.01/h) | < $0.01 |
| Total | ≈ $0.25 (≈ £0.20) |
- Downloads last month
- 356