[proposal] detect control decay as a red-to-green flip in a periodic ablation matrix, not as a per-PR trigger — measured preconditions, and the base-rate argument against building it

[os-project-manager]

Path: none | 仪器提案:把控制腐化当作周期性消融矩阵里的「红转绿」翻转来测 | ⚠️ 立卡席自己带了反对它的基率论证,⛔ 未主张该建
分诊重测与定级:2026-09-20T18:58Z
Filed by the domain:cli execution PM seat (#6024, session session_01QCdUBjM47SxioST9z5Zwdf) out of the #19371 feasibility round, whose verdict was that the per-PR trigger is not mechanisable. ⛔ Filed bare: finding only; domain:*, type and priority are triage's.

⚠️ This is a PROPOSAL with a measured argument on both sides. ⛔ It does not assert that the job should be built.

Dedupe words: ablation matrix red-to-green flip · control decay periodic job · neutral form per mechanism · mechanism register · #19371 reframe.

The reframe

#19371 asked whether "a PR adds a mechanism to a door" can be detected per-PR. Measured answer: no, at ≤12.5% precision and 8 fires/30 days on one package's one door family — details in that card's closeout.

⭐ But detection does not need the causing PR at all.

The decay is directly observable as a red-to-green flip in an ablation matrix: a control that used to go red when mechanism M was ablated, and now stays green.

Two properties make that different in kind from the per-PR trigger:

property	consequence
meaning-free at detection time	the semantic work moves to explaining a flip — which a human does once, on a real finding — ⛔ not to finding one
O(mechanisms), ⛔ not O(mechanisms²)	single ablation suffices. ⭐ Measured on the known instance: ablating the union fold alone already left the control 6/6 green. ⇒ the mechanism pair was needed to explain the cause and never to detect it

⛔ The argument AGAINST building it, stated first

• base rate: the [finding] two of three per-package pins had their lit control satisfied by an ECHO, not a real survivor — at least four more fixtures were never checked against the discriminant #18897 sweep ablated 21 fixtures across 30 legs and 9 mechanisms and found 1 instance.
• ongoing cost: one hand-authored neutral form per mechanism. ⚠️ These cannot be derived — the two mechanisms on the known door need opposite forms (authoringRuleUnionStack → identity, runPerPackageAuthoringRules → empty).
• convention gap: 222 of 340 packages/cli test files (65%) name no exported symbol in their leading docblock, and a further 47 name several ⇒ the matrix's rows cannot be populated mechanically from what exists today.

⇒ ⭐ a reader who concludes "not worth it" is reaching a supported conclusion, and that is a complete outcome for this card.

What would have to be true first

1. a mechanism register — what the ablatable mechanisms are;
2. one neutral form per mechanism, hand-authored and reviewed;
3. a decision on cadence, and on who reads a flip.

⭐ The cheap thing that already works, and its limit

PR #19369's shape: a mechanism that can reach a door someone else's control asserts on should OWN a marker in its output — there, the per-package pass prefixes where with package '<id>' — and the union run renders it bare.

⚠️ But its own docblock records the limit: that assertion is load-bearing only when paired with a positive control, because no real falsifier exists for that rule class. ⛔ So the marker convention is not a substitute for the matrix; it is what makes one control honest.

Related

#19371 (the feasibility verdict) · #18897 (the sweep that found the one instance) · PR #19369 (the repair, and the marker convention).

---

Generated by Claude Code

[os-project-manager]

Seat reading — held in pm:queue on purpose, with the reason

domain:cli execution PM seat #6024, patrol R84, read at 2026-09-20T19:57Z. A free dispatch slot existed at this tick and this card did not take it. Recording why, so the hold is a decision someone can overturn rather than drift.

This is not an implementation card, and its own text says so

Its three preconditions — "what would have to be true first" — are decisions, not work:

1. a mechanism register (what the ablatable mechanisms even are);
2. one neutral form per mechanism, hand-authored and reviewed — and the card measures that these cannot be derived, because the two mechanisms on the one known door need opposite forms (authoringRuleUnionStack → identity, runPerPackageAuthoringRules → empty);
3. a cadence, and who reads a flip.

None of the three is something a dev round can settle by measuring, and all three commit ongoing hand-authored maintenance. Dispatching against this card today produces one of two outcomes, both waste: a round that re-derives #19371's feasibility argument, or a round that builds a register and a set of neutral forms for a job the card itself argues against.

The card already carries its own verdict, and it is a legitimate one

⭐ It states outright: "a reader who concludes 'not worth it' is reaching a supported conclusion, and that is a complete outcome for this card." The measured base rate behind that — the #18897 sweep ablated 21 fixtures across 30 legs and 9 mechanisms and found 1 instance — is the strongest fact on the page, and it points away from building.

⛔ This seat is not recording that verdict as taken. The card is this seat's own output (filed out of the #19371 round it dispatched), and a seat grading its own proposal's worth is the conflict the protocol keeps away from. What this comment records is narrower and checkable: the card is not dispatchable as implementation work in its current shape, whoever ends up deciding its fate.

What would make it dispatchable

A ruling on precondition 3 alone would be enough to start: name the cadence and the reader, and preconditions 1 and 2 become a bounded, orderable first round scoped to the mechanisms of one door family rather than the repo. ⛔ Absent that, any fence this seat wrote would be a guess, and this seat has been wrong about a remedy's shape six times this shift.

The other reason the slot stayed open

Two rounds are in flight (#19413, #19424) and both were dispatched with an explicit invitation to refuse the order's sketch. #19422 is the precedent for what that produces — a measured disagreement that changes the deliverable — and a round that comes back that way needs seat capacity, not a third dev competing for it. Holding one slot against the patch rounds those two may need is the better use of it than a p3 proposal that needs a decision first.

⚠️ ⛔ No label moves with this comment: the card stays pm:queue, and its routing stays triage's.

---

Generated by Claude Code

[os-project-manager]

Dissent on the grade, ⛔ not on the card — pm:queue does not match what this card's own text asks for

_{domain:cli execution PM seat #6024 · session session_01QCdUBjM47SxioST9z5Zwdf · written 2026-09-22T03:00Z · dissent posted FIRST, pm:retriage applied immediately after, per the label-discipline rule}

The maintainer asked this lane to resume dispatching, p2 and p3 included. Every open domain:cli card was re-derived first-hand for that (⛔ none from a remembered verdict). This card is the lane's only unassigned pm:queue card, so it is the one the queue view offers — and reading it to its last line, ⛔ it is not dispatchable, and it is not marginally so.

Why the grade and the card disagree

pm:queue means 「有具名落点或复现的具体缺陷,或范围明确的工具/门禁修复,无可问之事」. This card has three open questions on its face, and they are its own 「What would have to be true first」 list:

1. a mechanism register — what the ablatable mechanisms even are;
2. one neutral form per mechanism, hand-authored and reviewed — and the card MEASURES that these cannot be derived, because the two mechanisms on the one known door need opposite forms (authoringRuleUnionStack → identity, runPerPackageAuthoringRules → empty);
3. a cadence, and who reads a flip.

⛔ None of the three is settleable by a dev measuring something, and all three commit ongoing hand-authored maintenance. That is the needs-user-decision shape as SKILL.md spells it — 「设计卡、feature/契约形状提案、需要 appetite 的多周程序」 — not the queue shape.

Stronger still, the card argues against itself and says so in as many words: 「a reader who concludes 'not worth it' is reaching a supported conclusion, and that is a complete outcome for this card」. Its own strongest measurement points the same way — the #18897 sweep ablated 21 fixtures across 30 legs and 9 mechanisms and found 1 instance.

⇒ dispatching against this grade produces one of exactly two rounds, both waste: one that re-derives #19371's feasibility argument, or one that builds a register and a set of neutral forms for a job the card itself argues should not be built.

What is being asked for, and what is ⛔ not

• Asked: a re-judge of the grade. The natural target is needs-user-decision, so that the build / don't-build question reaches the maintainer's inbox instead of sitting in a queue nobody can act on.
• ⛔ Not asked, and ⛔ not taken here: the verdict itself. This card is this seat's own output (filed out of the [finding] a control can be correct when written and decay into an impostor-satisfied one with NO edit to it or its fixture — measured on union-fold-command-parity, which stopped pinning the union fold when a sibling mechanism reached the same doors #19371 round this seat dispatched), so ⛔ this seat does not grade it — that is triage's, and raising a dissent is not grading. The original grade stays authoritative until triage re-judges it.
• ⛔ Not a complaint about triage's throughput. Triage re-measured and graded this card at 2026-09-20T18:58Z, ~1h46m after it was filed. The disagreement is about one card's shape, ⛔ nothing else.

Why it is raised now rather than left to drift

Seat reading 5752275493 (patrol R84) already recorded this card as held-in-queue-on-purpose, 「so the hold is a decision someone can overturn rather than drift」. ⚠️ That reading left the label saying the opposite of the reasoning, which means every seat's queue view — 取卡全序, the priority sort, any batch decision — still counts this as work to hand out. That is the same shape the half-state sweep files as H58 on other people's cards (pm:queue while the card's own text declares it is not queue work), and it should not need a sweep to find it here.

---

Generated by Claude Code