Promotion 2026-09-23 (2): staging → main

[modelmirror]

Promotion batch — 2026-09-23 (second)

Staging → main. This is a follow-up to promotion/2026-09-23 (#1977): the case-summary lane's accuracy and artifact fixes, found in its two staging rehearsals, plus CI and docs hygiene. No process digest moves: no prompt the registry hashes (predict.md, evaluate.md), no predictor or evaluator config, no DEFAULT_MODELS, ENGINE_RETRIEVAL or MCP manifest entry. No cell's execution path changes.

PR	What	Class
#1979	SECURITY.md / docs/security.md: the review environment admits main and staging (so summarize's staging rehearsal can pass its hold); drops a stale dedicated-key clause	security docs
#1980	Gate coverage measured under sys.monitoring (COVERAGE_CORE=sysmon): the test step roughly halves, with an identical coverage total	scripts/gate.sh, docs
#1981	summarize.md: an "Accurate to the record" block covering attributed allegations, a question's direction, procedure in the docket's own terms, and dates and counts as shown	summaries prompt, docs
#1982	summarize.md: a filed request is not reported as granted	summaries prompt, docs
#1983	summarize.yml stage job: an allowlist check of the staged records before the public summary-stage upload	workflow, summaries

Engine smoke is waived (promote:skip-engine-smoke): nothing here reaches a cell, an engine CLI, an engine action or the invocation path. The evidence is a green scenario=all-offline run at the promoted head.

Effect checks (report each as it lands)

1. Summaries backfill, with the fixed prompt and the stage check. Release the next summarize run's review hold on main (03:43 UTC, or a dispatch). Expected:
   • the plan reports about 195 cases, estimated at about $15–25;
   • the stage job summary reads stage check: N staged case record(s) kept, 0 removed, where N is the number staged;
   • generate reports about 195 written;
   • each summary's front matter carries the new prompt digest (not sha256:91da8a2b… or sha256:7500b836…);
   • a summaries/refresh PR opens.

   gh run list --workflow summarize.yml --branch main --limit 3
   gh pr list --head summaries/refresh --state open --json number,title,url
   

   Spot-check a sample of that PR against its filings before merging it (summaries: plain-language case summaries per snapshot #1968), looking in particular for any request reported as granted.
2. Faster gate (ci(gate): measure coverage under sys.monitoring #1980). The next push main CI run's gate → test step takes about 200–260s, down from about 450–500s:

   gh run list --workflow ci.yml --branch main --limit 1 --json databaseId --jq '.[0].databaseId'
   gh run view <id> --json jobs --jq '.jobs[]|select(.name=="gate")|.steps[]|select(.name=="test")|((.completedAt|fromdateiso8601)-(.startedAt|fromdateiso8601))'
   

3. Docs (docs(security): record the review environment's deployment branches #1979). No run-time observable.

🤖 Generated with Claude Code

[modelmirror]

Tagged promotion/2026-09-23-2 (annotated) on merge commit dfed43e. Effect checks pending: the summarize backfill (next run's hold on main; its stage summary should read … kept, 0 removed) and the faster gate on the next push main CI run — each reported here as it lands.

[modelmirror]

Effect check 1 landed: summarize run 35921436250 on main planned 195, stage check: 195 staged case record(s) kept, 0 removed, wrote 194 ($18.77; one rejected for markup) under prompt digest sha256:114c7d00…, and opened #1986. After a spot check (#1968) five summaries misstating a held petition's posture were withdrawn; #1986 merged with 189 summaries (16cb79e) — git ls-tree -r --name-only origin/main data/cases | grep -c /summaries/ → 189. The remaining six cases are owed and wait on #1987's promotion.