H1 on likelihood: check 5 passes at half and a quarter, and OpenStack carries a contrast¶

# research(2026-10). Read under the rules fixed in "Planned before any likelihood result is read" above, committed (ea562bf) before the 76 likelihood jobs were pulled. likelihood-partition-*-rp1.0.json, likelihood-sesoi.json (74e8195, computed by rule 3 as fixed in ea562bf, which leaves no free parameter; the job logs committed with it print each run's arm means), partition-pilot-<org>-likelihood.json and -likelihood-total.json.

The SESOI carried from exact match by the positive control: adapters raise the per-token log-probability of their own half's references over the base model by 0.4372 nats and exact match by 0.2384, over 92 arms in 46 runs, so SESOI = 0.01 * 0.4372 / 0.2384 = 0.0183 nats per token.

Check 5 on likelihood passes. Planted in half of each half's refinements: OpenStack 10 of 10, Wikimedia 10 of 10 (greedy exact match: 0 of 10 and 2 of 10). Planted in a quarter, below what check 5 requires: OpenStack 10 of 10. The measure sees a convention learned at a minority rate.

Greedy exact match planted at 0.75 (rq1-partition-*-plant0.75.json, partition_run.sbatch PLANT=0.75, 20 runs, the measurement the check-5 entry queued for option 1): every run's own verdict is pass (own minus sibling near +0.2), and every run fails check 5 on test 6, because a sibling adapter that writes its own marker on nearly every output never matches the other half's references exactly, so its exact match is 0, which test 6 reads as degenerate. Greedy exact match therefore detects a convention applied in three refinements of four and not one applied in half; likelihood, scored at half and a quarter, detects both.

H1 on the development window, likelihood (22 runs an organization, own minus sibling in nats per token, the registered Holm levels; total log-probability, the registered sensitivity, and greedy exact match, the registered secondary, beside it at 97.5%):

OpenStack's adapters assign their own half's references more probability than the sibling half's adapters do, at both levels; the estimate is above the SESOI and the lower bounds are below it. Wikimedia's contrast is positive at 95% and sits inside the SESOI band at both levels. Total log-probability agrees in direction on both. Exact match reads both as inconclusive, as before. K, sized on the likelihood pilot's run spread at the SESOI, falls to the floor of 10 on both. This is the development pilot, not a confirmatory read; the confirmatory cells read the sealed test window.

Still to do before Stage 1: the H1 sensitivity simulation on likelihood (partition_sensitivity.py simulates exact match from the placebo run; it needs the metric), and the manuscript's H1 measure, SESOI, K and check 5 rewritten from these readings.