拓展阅读 / NanoJev：小型决策模型

知识库收录材料 · 原始文件 ↗ · 来源、版本与许可

Fable implementation assistance and code review

On 2026-09-17, two local Claude Code requests explicitly selected claude-fable-5-1 with --effort xhigh. Both returned that canonical model identifier in their usage records. Tools, Chrome integration, customizations, and session persistence were disabled. Only an authored specification or selected project source was supplied; no credentials were included.

These are the CLI's reported list costs, not a reconciled Vercel invoice. The first request demonstrates that the CLI budget option is checked after requests and is not a strict per-request billing ceiling. The second request allowed one turn and returned its review within the remaining allocation. Private prompts and raw response logs remain outside the repository.

What was reviewed

The successful review received complete snapshots of snake_game.py, test_snake_game.py, and calibrated_objectives.py, plus the training integration diff for train_pipeline_decisions.py.

Fable approved the core Snake transition rules: reverse moves are rejected; nongrowing moves may enter the cell vacated by the tail; eating retains the tail; food never occupies the body; full-board completion is a win; collision preserves the last valid body; and transitions do not mutate their input. The stored RNG state preserves reproducible continuation through JSON round trips. Rendered model inputs contain physical state and explicit action-specific propositions, with no RNG or exact safety labels.

The review also approved the paired proper-reward gradient, its detached baseline, complete-question normalization, Boolean [0,z] representation, and the common optimizer-update denominator. This is a review of our explicit research implementation. It does not establish TypeSafe's undisclosed training procedure or demonstrate a learned model's performance.

Probability-gradient conclusion

For independent categorical draws A_i from p, with replacement, the implemented reward is:

R = (2/M) sum_i 1[A_i=Y]
    - sum_{i!=j} 1[A_i=A_j] / (M(M-1))

For each score-function term, retain the reward terms that depend on that sample:

h_i = (2/M) 1[A_i=Y]
      - 2 sum_{j!=i} 1[A_i=A_j] / (M(M-1))

gradient E[R] = sum_i E[h_i gradient log p(A_i)]

The factor of two accounts for both ordered-pair positions. This gradient identity does not require or assert R = sum_i h_i.

The baseline for sample i depends on the other samples, the observed outcome, and detached probabilities, but not on sample i. Its expected score-function contribution is therefore zero. This requires conditionally independent model draws and an outcome observation independent of those prediction draws under the specified event experiment. It applies for every finite M >= 2. No group-standard-deviation normalization is used.

Unbiasedness does not guarantee lower gradient variance for every distribution. We did not adopt the review's suggestion to require the conditional-mean baseline to reduce gradient variance in every test. Numerical gradient checks and event-adapter validation are separate from this text review.

Findings and follow-up decisions

Changes after the paid review

The reviewed action target was uniform over collision-free moves. At the user's request, the final action question now uses a local food-progress heuristic: first avoid one-step collision, then minimize Manhattan distance from the next head cell to the current food, sharing probability uniformly across ties. If every offered move collides, all offered actions tie. This is an action-policy target, separate from the individual safety propositions and from global Snake planning.

That change, the seed-grouping correction, and the additional tests were implemented and checked by the project agent after the paid review; they were not sent for another paid review. The source-group description was also corrected: this module groups an episode by size and canonical seed and does not claim to perform symmetry grouping.

Seven standard-library tests pass, covering determinism, JSON continuation, immutability, reverse and terminal rules, wall/body collisions, tail entry, growth, a full-board win, hidden-input exclusion, schema compatibility, food progress, tied moves, and the all-dangerous-action case:

PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s scripts -p test_snake_game.py -v

An additional comparison against the exact reviewed source checked 2,880 states across board sizes 8, 16, 32, and 50. Safety-event input bytes, event specifications and outcomes, transition/RNG results, and ordinary nonnegative seed groups were unchanged. A preceding 8,640-candidate-transition check covered state immutability, safety/outcome consistency, and training-row validation. These are environment and interface checks, not evidence of a trained policy's playing strength.

快照未包含部分工程附件或配图；请通过上游链接或原始材料查看，收录范围见来源清单。