v9_setb_confirm_golden.md

v9 Set B confirmatory host golden

Generated before any board confirmatory run for the frozen v9 variant.

• Frozen variant: pbil_island8_graded_v9
• Confirmatory seed: 0xB17D
• Variant arm seed: 0xB17D ^ 0x4A4A = 0xFB37
• Random arm seed: 0xB17D ^ 0xBEEF = 0x0F92
• Train frames per condition: 64
• Final holdout frames per condition: 256
• Nominal host-screening arm budget: 22387 candidates
• Primary confirmatory metric: hard holdout delta

Golden at budget 22387

pbil_island8_graded_v9 raw 0x7c009161b2 phase 18 threshold 13 majority_idx 3 taps 0xf80122 train 83489 83968 holdout 128 1024 evals 1879791616
graded_holdout 382336 385024

random raw 0x06c219610c phase 12 threshold 8 majority_idx 3 taps 0x0d8432 train 38 256 holdout 15 1024 evals 5731072
graded_holdout 332459 385024

Hard holdout delta:

128 - 15 = +113 / 1024

This exceeds the preregistered Set B confirmatory threshold +8 / 1024.

Evals-field note

The C twin's variant-graded CLI reports evals for graded-search variants as the accumulated graded train bit-count (1879791616 above). The board mailbox uses the project-wide board convention: evals means fabric frame evaluations. For the nominal Set B budget, the expected board frame-eval count per arm is:

22387 candidates * 4 train conditions * 64 frames = 5731072 frame-evals

The board checker therefore validates the variant mailbox evals field against budget * 256, while validating genome, train score, hard holdout, and graded holdout against the C twin golden.

Budget derivation

The board confirmatory firmware must run its graded-path speed probe and derive the actual board budget from measured evals/sec. If the measured board budget differs from 22387, host/check_v9_confirm_mailbox.py recomputes the frozen variant and random goldens for the observed budget before comparing mailbox payloads. The variant and seed remain fixed; only the preregistered speed-derived budget may vary.