[morgaesis]

Hosted reviews move to openai/gpt-5.6-luna, pinned to the Azure zero-retention endpoint with a price ceiling of $0.22 and $1.32 per million input and output tokens. The change is worth making for one number: gate-verdict correctness, how often the merge gate blocks exactly when it should, goes from 73.7% to 85.5% across repeated runs of the 70-fixture suite. That is the metric that decides whether a review costs somebody a merge, and it has been our weakest for months.

Seven scored runs support it. No run produced a false finding and every run stayed silent on all thirteen clean fixtures. Detection rate is level at 89.5%, because the suite cannot resolve a difference smaller than its own run-to-run spread; it is reported rather than relied on. Reviews cost roughly a seventh as much, and ninety-fifth-percentile latency falls from 29.1 to 17.6 seconds, which matters because the per-model slot a hosted review runs under has to be sized for that tail.

Three of the seven runs enforce the contract the deployment actually uses: one pinned upstream provider, zero data retention, and the manifest price ceiling. That distinction turned out to be load-bearing. The other finalist looked steadier on unconstrained screening and lost its entire latency advantage once pinned, because the endpoint the router had been choosing for it is not one a zero-retention deployment may use. The recorded baseline therefore comes from a contract-enforced run, and it is the median on detection and the lowest of the three on gate correctness, so the release floor sits inside the observed spread instead of on top of it.

conservative_context_tokens gains the new generator; without that entry its review budget would silently drop to the unknown-model floor.