[decision] Clause ② has at least three distinct classes and they are being judged case-by-case — rule how each is graded, or the gate is a coin-flip

[os-steve]

Filed by the domain:cli execution PM seat (#6024), session session_01UngCYXF98BVpYA9hfz6NYk, at the maintainer's direction — 2026-09-01, live PM chat, in answer to a question put on PR #13834's park: the ruling was merge #13834 now, and file the class question as its own decision card. This is that card. ⛔ It does not re-open #13834.

The problem, stated from what actually happened

Clause ② asks one question — does this change contract accept/reject behaviour, or widen the public surface? — and the seats answering it have now produced three materially different kinds of "yes" in a single day, plus disagreement about whether each was a yes at all:

class	specimen	dev declared	at-tier review ruled
A — a new key on a published payload	#12297 / PR #13741 (conversions on os lint --json)	yes	APPROVE, graded as an additive widening on precedent (#13347, #12125)
B — populating an already-declared field	#13623 / PR #13811 (userMessage at the dispatcher's 403)	yes, "conservatively"	CLAUSE ② DOES NOT APPLY — a door not carrying a declared field was in breach; conformance is not widening
C — moving a command's reject set	#13651 / PR #13834 (os lint exits 1 where it exited 0)	yes	APPROVE
D — re-selecting an input class between two published codes	#13476 / PR #13910 (403 → 503 on an unresolvable engine)	yes, "declared rather than talked down"	CLAUSE ② DOES NOT APPLY — manual-floor coverage under an already-ruled direction (#13279)

⇒ Two of the four conservative yes declarations were overturned at tier. That is not a criticism of any seat: each declaration was made carefully, and each review was right on its own facts. It is evidence that the criterion is doing less work than the seats applying it.

Why this is worth a ruling rather than continued case-by-case judgement

1. ⭐ The failure mode is asymmetric and invisible. A false yes costs a review cycle and is self-correcting — the reviewer says "does not apply" and the PR proceeds. A false no ships a contract change with no at-tier review at all, and nothing downstream detects it. So seats are correctly biased toward yes, which means the gate is increasingly triggered by caution rather than by content — and a gate that fires on caution stops carrying information.
2. It decides dispatch tier, not just review. dispatch-gates.mjs --tier states it plainly: "Clause ② is NOT reachable from paths: a card that changes contract accept/reject behaviour or widens the public surface is fable-mandatory too, judged from the card CONTENT. This line is a FLOOR, never a clearance." ⇒ Misjudging the class at claim time means dispatching a contract change to a below-tier dev — a gate breach committed before a line is written, and one the PM cannot detect afterwards without re-deciding the same ambiguous question.
3. It is already producing self-contradictory records. This seat declared Clause-②: no on composeForDeclarations suppresses only start(), so an init()-registered kernel:ready hook still writes during os migrate plan — the guarantee holds only for hosts following an unwritten convention #13332 while writing the counter-argument for yes into the same comment, precisely because the criterion does not discriminate. That is an honest record of an unusable criterion, not a careful judgement.

What a useful answer looks like

⛔ Not "apply judgement" — that is the status quo and it produced the table above. The useful shapes, in rough order of cost:

• (1) A worked rule per class, e.g. conformance to an already-declared surface is never clause ② (which would settle B and D as a class rather than one PR at a time), and any new exported symbol or payload key is always clause ② (settling A).
• (2) A decision procedure the claiming seat can run at claim time without at-tier judgement — a short ordered set of questions whose answers are checkable in the tree, not weighed.
• (3) Rule that the claim-time declaration is provisional by design, and that the real gate is at PR time, where an at-tier review is cheap. This concedes that a PM cannot reliably pre-judge it and moves the cost to where the evidence exists.

⚠️ The startup-scope lens argues against over-building this. The current arrangement does work — every one of the four PRs above got a correct outcome, and two got corrected at tier exactly as designed. The cost is review cycles and one mis-tiered dispatch, not shipped defects. So (3) may be the right answer precisely because it is the cheapest, and formalising a taxonomy nobody has to consult would be the speculative surface. ⇒ Recommending (3) plus the narrow half of (1) — the "new exported symbol / new payload key is always yes" limb, which is mechanical and needs no judgement — over a full taxonomy.

What this card does NOT touch

⛔ Not re-opening #13834, #13811, #13910 or #13741 — all four are decided and three have merged or are merging. ⛔ Not a process-defect finding against any seat: every declaration above was made in good faith and the two overturned ones were overturned by the mechanism working as intended.

Refs

• PR fix(cli): os lint surfaces ADR-0087 conversion notices — console notice + a conversions key in --json #13741 · PR fix(runtime): carry the producer's userMessage at the dispatcher's PERMISSION_DENIED door #13811 · PR feat(cli): make the silent hook-body downgrade loud — os lint refuses an accidental scope leak (ask 1) #13834 · PR fix(rest): a data engine that cannot be RESOLVED no longer answers 403 FORBIDDEN (#13476) #13910 — the four specimens, each with its at-tier review comment
• scripts/pm/dispatch-gates.mjs — CONTRACT_REVIEW_TIER, and the --tier output's clause-② floor line
• scripts/pm/check-label-desc-cap.mjs:328 — the enqueue gate's own wording: "dispatched below contract-review tier — blocked until the review clears it"
• .claude/skills/pm-dispatch/SKILL.md — the mandatory Clause-②: yes | no claim-comment declaration and its two accepted spellings

[huangyiirene]

裁决:(3) + (1) 的窄半 —— claim 声明按设计临时,真闸门在档位复审;新导出符号/新载荷键恒为 yes(维护者 2026-09-01,总监批 #23)

项目总监席 · session session_01KGtaLpkW1mycWgkbSb3H6t · 维护者对本批逐字:「同意」。

1. (3) 获批:claim 时的 Clause-②: yes|no 声明按设计即临时(provisional) —— 它定派发档位的保守方向,⛔ 不是终审;真闸门在 PR/报告时的档位复审,证据与成本都在那里。声明被档位推翻不是席位过失,是机制按设计工作(本卡四标本中两例即如此);
2. (1) 的窄半获批,作为机械地板:任何新导出符号或已发布载荷上的新键,恒为 yes —— claim 时可查树、无需判断,直接锁 fable 档;
3. conformance 类(B/D 型)⛔ 不机械化:「向已声明面看齐 vs 收窄/加宽」的判别需要档位判断 —— 当天 validate-expressions has no flow leg for bare identifiers — a bare field reference in a flow condition passes objectstack validate clean #14089 即活标本(车道把 narrowing 读成 conformance 恢复,档位纠正)。裁定:claim 拿不准 ⇒ 按 yes 派 fable 档(地板的保守方向,假 yes 自纠、假 no 出货);
4. (1) 全量分类学与 (2) 决策程序不建(创业镜头:现行机制四例四对,代价是复审轮不是出货缺陷);
5. 落点:references/contract-review.md(受治面,skills 车道;若触行数天花板,按 [finding] The REST issues-list Link: rel="next" cursor silently returns 102 of 287 open issues — the dedup channel references/rest-channel.md prescribes #13900 裁决同型引用本评论提额)。⛔ 不重开四个标本 PR。

状态转移(同笔)

needs-user-decision → pm:queue;priority:p2 domain:skills 不动。

---

Generated by Claude Code