Datasets:
Appendix B.1 — the licence suite, check by check
One row per check. Reference pretrain_iter_0484865. The suite has nine checks, not eight: check8_nulls__suite_7 in the stored artefact holds two — 7a scrambled and 7b split-half — which is the whole of the §4.1 / §2.6 discrepancy. Eight stored keys, nine checks.
| # | check | pass criterion | outcome | ran on |
|---|---|---|---|---|
| 1 | twin agreement | none — a characterisation | the two batteries' profiles agree: median cosine 0.995 for For |
all 188 (188 of 188) |
| 2 | convergence | none — a characterisation | the profile reaches cosine 0.99 with its full set by a median 56 of 720 prompts | 31 per-prompt (31 of 188) |
| 3 | length | none — a characterisation | three padding levels; min pair cosine and the length index | 31 per-prompt (31 of 188) |
| 4 | re-weighting (ν) | none — a characterisation | a second generator menu against the twin floor | all 188 (188 of 188) |
| 5 | calibration | none — reported as inflation | do the stated intervals cover the twin difference at the nominal 95% | all 188 (188 of 188) |
| 6 | companion | none — a measured cap | the measured version of the pre-run companion cap | 8 with companions (8 of 188) |
| 7a | scrambled null | PASS: the real norm must exceed the null's 95th percentile | median 10.7× the null p95, minimum 7.2×; 0 of 31 at or below | 31 per-prompt (31 of 188) |
| 7b | split-half null | PASS: the real norm must exceed the null's 95th percentile | median 12.8× the null p95, minimum 6.0×; 0 of 31 at or below | 31 per-prompt (31 of 188) |
| 8 | σ₀ floor | none — a sensitivity | how far the smooth floor's open parameter moves the reading, in twin units | all 188 (188 of 188) |
Only two checks have a pass criterion — the two nulls — and both pass for every construct they ran on. The other seven are characterisations or sensitivities: they are reported, and there is nothing for them to fail.
Four checks were re-run after the per-prompt join fault of 19 October: 2 convergence, 3 length, 7a null_scrambled, 7b null_split_half. Table B.2 gives published against corrected per construct.
Understated for calibration. Check 5 reports an inflation factor rather than a pass: the stated intervals cover the twin difference at below their nominal 95% rate, so the intervals are optimistic and the inflation is the amount by which. It ran on all 188 constructs.
Why four checks ran on 31 and one on 8. Checks 2, 3, 7a and 7b need per-prompt vectors, and the record keeps those for 31 constructs; the other 157 are profile_source: accumulators. Check 6 applies only to the eight constructs that can have a companion (ClassDef, For, FunctionDef, If, Match, Try, While, With). Neither is a gap in the run; both are properties of what the record holds.