CMA-ES over the eighteen fair-leaf weights inside the fast depth-4 five-stratum search, fresh training seeds per generation, one held-out paired screen

Primary metric

paired mean whole-game score delta, candidate minus frozen leaf, depth 4 five strata, 64 held-out games

Statistical unit: whole-game

Pass criteria, fixed in advance

• All CHECK gates passed before the first training seed was read: leaf bit-identity (0 mismatches), search parity with FastSearch (0 action/work/node/cache/depth mismatches), determinism across 1 and N threads, reflection, metadata blindness, completed depth and legality under perturbed weights.
• Every generation artifact and the screen artifact have incompleteDecisionsTotal 0, illegalDecisionsTotal 0 and no score-identity failures.
• Held-out screen, depth 4 five strata, 64 paired games on 0xa52b0000-0xa52b003f: bootstrap 95% lower bound of the paired score delta > 0 AND Student-t 95% lower bound > 0.
• Held-out screen: candidate lower-quartile score >= frozen lower-quartile score.
• Held-out screen: paired mean score delta > 0 in both halves (seeds 0-31 and 32-63).
• The candidate evaluated is the CMA distribution mean written to candidate-weights.txt before the held-out lease is opened; no other vector is screened.

On pass: Freeze candidate-weights.txt with its SHA-256; register a fresh-block replication by a different runner and a seven-stratum-native evolution as successor experiments; do not open protected or final seeds.

On fail: Record a valid run with scientific outcome fail for this exact configuration (optimiser, budget, depth, strata, fitness); do not adopt the vector; report the 7-stratum transfer arm as diagnostic only; open no further cohort for this candidate.