cold-seat-economy — re-audit on substance

Census id: F8 · Family: F — rules corpus + generated-rule tooling + shipped audit scripts · Provenance: framework pin b069c59328aa8e08671d3a19fd25efa618dfc1c5 (staging FETCH_HEAD, read 2026-09-11; re-anchored from census pin a1337cb301). Every file:line below is at that pin. Raw reference drafted by stage E4 — factual, unpolished by design.

What it bans

Spending full cold re-audits because a SHA moved while the judged substance did not, and burning the expensive seat on a non-final diff. .claude/rules/cold-seat-economy.md:1 «# Cold-seat economy — re-audit on substance, follow up with a fresh narrow seat»; the core: .claude/rules/cold-seat-economy.md:25 «A cold seat's verdict is re-earned by a change in what that seat judges, never by the head» / «commit moving. A SHA mismatch alone is a bookkeeping problem»; seat ordering: :53 «## §2 Seat ordering — the expensive WHAT-audit runs on the FINAL diff» with :55 «Dispatch cold seats serially, cheapest-consequence first, and run the expensive» / «WHAT-conformance audit only once the diff is final.» Class header: :6 «> Class: C — prose-only. The two calls this rule governs — «did the substance this seat judges change?» and «resume or fresh?» — are judgment, not mechanically detectable, so no gate is reachable».

Never (fires)

The named anti-patterns: .claude/rules/cold-seat-economy.md:105 «- #reaudit-on-sha-move — spending a full cold re-audit because Audited-SHA no longer prefixes HEAD, without asking whether the seat's substance moved.»; :111 «- #self-issued-verdict — the dispatching session writing the seat's verdict itself to save the re-audit cost. Counter: §1 third bullet — smaller cold check, never no cold check.»; :113 «- #expensive-seat-on-nonfinal-diff — dispatching the costliest audit in parallel with a seat whose findings can still change the diff.»; :115 «- #resume-sold-as-savings — justifying a resumed seat by token economy. The measurement says ~11%; resuming is the expensive continuity vehicle.» The hard floor: :48 «- Never self-issue the verdict to dodge the cost. … The cheap path is a smaller cold check, never no cold check.»

Always (clean)

A scope check then a narrow cold delta: .claude/rules/cold-seat-economy.md:31 «- Same deliverables, same permitted files, same descopes → the fidelity verdict stands in substance. Confirm scope mechanically (git diff --name-only <audited>..HEAD against the kickoff's permitted-file list), then re-establish the verdict with a narrow cold delta check: hand a fresh cold agent only the incremental diff, the kickoff's scope sections, and the prior round's watch-list». And the measured default — inlined inputs, zero tool turns: :79 «| fresh agent (executor tier), inputs inlined, zero tools | 85,855 | 0 | 16 s |» vs :76 «| fresh agent (top tier), full fidelity audit | 185,239 | 19 | 174 s |».

Enforcement channels

• Rule-index row: .claude/rules/00-rule-index.md:16 «| cold-seat-economy.md | C | re-running a cold seat on already-judged work; resume-vs-fresh choice. | skill-embed(2) |»
• Skill-embed, two hosts: .claude/rules/cold-seat-economy.md:3 «» and :4 «» — «Delivery = skill-embed at the two choreography owners (markers above); this file is the SSOT read on demand.» (:9-10)
• No CI mechanism at all (stated at :137 — «no CI mechanism at all; seats are session-read agents»); not backend-rendered, so no FF diagnostic applies (FF7001/FF7002, packages/core/diagnostics/registry.ts:321/:329). Honest status: two skill-embeds + review-time judgment; promotion path to a SCOPE-UNCHANGED/SCOPE-MOVED helper is scoped in §5 (:124-128) and deliberately not built.