grid-policy-mlp-repaired-v4

PPO-trained GridPolicyMLP after analytic parameter repairs from KimiBrainSurgery. Architecture: 500 → 32 → 32 → 5, tanh, 17,253 parameters. W2 solvers use analytic derivatives and local linearization; these repairs do not run gradient-descent training. Gates measure finite sets, not universal safety.

Frozen paired evaluation — greedy actions

Games	Seeds	Original mean	v4 mean	Paired gain	Wins / losses / ties
1,000	30000–30999	8.886	12.185	+3.299	253 / 20 / 727

v1/v3 use the same old set; v1's 1,000-game result includes its earlier 64-game pilot. Seventeen v3 failures informed v4 development. v4's old-set mean is 12.182 versus original 8.694; its separate fresh-set result appears above for v4. Zero observed v1 losses is not a never-worse guarantee. v3 and v4 have higher measured means but observed regressions, including the ascent pre-positioning class. These numbers do not estimate sampled-action play or the separate Qwen experiment.

The original v1 loop kept three of four edits because mean return on a reused 64-game gate did not decrease. Only one kept edit reports its targeted state fixed.

Current repair status

Experiment 8 tested 18 batch configurations against seven failure states under v4. None fixed all seven while preserving gate mean; no v5 was promoted. A candidate fixed one state at equal gate mean, but failed that all-seven requirement. This is a negative result for tested solvers, not proof that a larger network is necessary. Some former v4 fresh-set loss seeds now enter batch mining: future candidates need a new untouched final audit. The historical frozen v4 evaluation is unchanged. Batch evidence and limits.

Checkpoint and reproduction

policy_weights.npz contains float32 W1, b1, W2, b2, W3, b3. File SHA-256: 87c2248300f93f1bb395f0466602d3616edcf6125ea513a15803d5a84ef3ea40. Documentation synchronization did not alter these weights. Current evidence and status · Per-seed dataset · Live visualizer.

Independent work inspired by Sparks of Weight-Space Programming. Different policy provenance and repair protocols must be evaluated separately.