The pull-request workflow has no pre-open AC re-anchor, so a long implementation opens against a remembered ticket

Context

Authors do check ACs at intake. The gap is that the intake read decays across a long or complex implementation and nothing re-anchors it at PR-open (@tobiu: "it can happen to get 'lost in the forest', so a when-done check matters"). This is not a missing first check; it is a missing last one.

pull-request-workflow.md:160 forbids author self-review — correct, cross-model review is the gate. A pre-open AC re-anchor substitutes for nothing and is a different act.

Measured over the last 50 PRs (open + closed)

Sampled 8 RC bodies across families. Every one has the same skeleton — premise sound · placement right · not a D+S · but the implementation does not deliver what the ticket required:

• #16444 — "does not yet enforce #16443 … the guard is not invoked by CI"
• #16442 — "additive beside, rather than authoritative over, the existing verdict"
• #16466 — "can only detect churn when tests hand-inject history"
• #16421 — "do not yet deliver the requested behavior"
• #16436 — "the ticket explicitly required classification"

3 of 8 are one sub-pattern: the capability was built and the production path never uses it. Each is one command to falsify.

Economics: unit CI ~13m + a fanned-out review ~20m ≈ 33m per RC cycle; 37 cycles ≈ 20 hours in this window, before counting D+S, which costs the implementation rather than a cycle. Operator bar: "if a self-AC check reduced 20% or more RCs => already worth it" ≈ 4 hours plus GPT tokens, currently the binding constraint.

The Fix

Two questions in the pull-request workflow's pre-open sequence. Budget-verified: 197 bytes, against skills.manifest.json:14 maxPositiveDeltaBytes: 250 (gates per-file and aggregate):

**Pre-open AC re-anchor:** re-read the LIVE ticket. Per AC saying enforced/invoked/authoritative: name the production-path observable and run it. A test injecting state proves the unit, not the AC.

Question 1 targets the D+S class — #16457's D+S was exactly a stale-ticket-premise failure: I had amended that body myself, then opened against my memory of it. Question 2 targets the 3-of-8 wiring class.

Contract Ledger

Acceptance Criteria

• The insertion lands in the pre-open sequence, not in a review or intake section, and does not read as authorizing self-review.
• npm run ai:lint-skill-manifest passes; the net positive delta is ≤ 250 bytes, measured and stated in the PR.
• No new file, no new section heading, no §ref — a pointer-free inline addition.
• The wording names production-path observables explicitly, since "a test injecting the state" is the exact defect 3 of 8 sampled RCs describe.

Out of Scope

• Any mechanical enforcement. A hook that re-reads a ticket and diffs it against a diff is a much larger design; this is discipline-only text.
• Author self-review — :160's ban stands untouched.
• Review-load redistribution. The 76% GPT concentration is the larger lever and is a separate, operator-owned decision.
• Reviewer-rigor questions. I probed review body size as a rigor proxy and it failed: Claude reviews are the longest (mine the longest of all, 13.5k median) and Emmy's the shortest at 8k while finding more. Under a mandated template, length is fixed overhead and shorter can mean better. No claim about reviewer rigor is carried here; I have no instrument for it.

Avoided Traps

• Do not target "PRs merged without an RC." That metric pushes toward smaller PRs and pre-negotiated reviews; some RCs are the system working (Emmy's D+S on #16457 produced a better decomposition than my own reduction chain). The quantity to reduce is author-findable findings per PR.
• Do not add prose. The 250-byte cap is the feature — it forces two questions instead of a checklist nobody re-reads.
• Do not present this as a review substitute. It removes what an author can reach so the cross-family cycle spends itself on what they cannot.

Related

• #16448 / D#15758 — unrelated lane; cited only because #16457's D+S is the worked example above.
• #16457, #16405, #16400, #16383 — the four D+S in the sample. #16444, #16442, #16466, #16421, #16436, #16465, #16416, #16415 — the sampled RC bodies.
• .agents/skills/skills.manifest.json — the byte budget. ai/scripts/lint/lint-skill-manifest.mjs — its enforcer.

Live sweep at creation: no open ticket proposes a pre-open AC check; the nearest substrate is pull-request-workflow.md:160's self-review ban, which this deliberately does not touch. A2A sweep: no [lane-claim] on the pull-request skill.

Origin Session ID: 11695cce-9854-4be2-80c3-8ea4322298bf

Retrieval Hint: query_raw_memories("pre-open AC re-anchor pull request workflow 79 percent RC rate production path observable")

Authored by Vega (Claude Opus 5, Claude Code) — from @tobiu's friction→gold framing and a 50-PR measurement.