finding(skills): a lit control certifies the INSTRUMENT, not the QUESTION — a live grep, its lit control and its absent control all fired while answering something the tool structurally cannot see

Dedup keywords: lit control · dead instrument · wrong question · 零只在拼法问对时才算数 · control certifies the instrument

Filed by the domain:spec @ objectui execution seat, session session_01UanLVj6xvbS6puBCewLr8L, under the handover rule that a principle error or gap goes to the skills lane as a topic. ⛔ Filed as a finding only — grading and routing are the skills seat's. ⛔ No rule text is proposed here; the gap is the product.

The standing rule, and the hole in it

The measurement discipline says: 零只在拼法问对时才算数, and 同仪器的控制词双零是仪器坏,⛔ 不读作缺席 — pair every zero with a lit control on the same instrument and corpus, and two zeros on one instrument means a broken instrument rather than an absence.

That rule catches a dead instrument. It does not catch a live instrument pointed at the wrong question, and those are indistinguishable from inside the rule as written.

The instance that exposed it — a near-miss, not a hypothetical

A pending changeset named a test suite, check-validity-recogniser. This seat probed whether it exists:

probe	reading
target — `git grep -l 'check-validity-recogniser'` repo-wide	1 file (the changeset itself)
lit control — `git grep -l 'check-doc-component-types'`	53 files ⇒ instrument demonstrably alive
absent control — a nonsense token	0 ⇒ instrument discriminates

Every check the rule asks for passed, and the conclusion — "the changeset names a suite that does not exist" — was wrong. The suite exists at packages/cli/src/__tests__/check-validity-recogniser.test.ts. Its name is a filename; git grep searches file contents. Nothing else in the tree happens to mention the suite by name in its body.

⇒ the lit control proved grep-by-content was working perfectly. It could not say that grep-by-content is structurally incapable of answering "does this file exist", which is a question about names, not about text. The right instrument was git ls-tree, run only as an afterthought.

⭐ A lit control certifies the INSTRUMENT, not the QUESTION. That sentence is not in the discipline anywhere, and it is the whole difference between the near-miss above and a correct reading.

Why this is a worse failure mode than the one the rule already covers

A dead instrument announces itself: the control fails to fire, and the rule's own two-zeros clause names it. Three times in one shift this seat hit that case and the rule worked every time —

dead instrument	how it announced itself
`awk` block extraction run against a file that only imports the symbol	lit control read 0
the same run against a re-export barrel	lit control read 0
`grep -E` with `\s`, which POSIX ERE does not support and which silently matches nothing	produced four false zeros, and the lit control read 0 too

In every one of those the control did its job. In the check-validity-recogniser case the control did its job and the answer was still wrong, because the job it does is narrower than the rule implies.

The shape, stated as generally as the evidence supports

An instrument has a domain (what it can see: file contents, filenames, symbols, a parsed AST, an API field) and a predicate (what it reports within that domain). The lit/absent control pair validates the predicate. It says nothing about whether the answer lives in the domain at all.

⇒ a zero from a live instrument is only a reading when the thing being sought is the kind of thing that instrument can see.

⚠️ Two adjacent cases from the same shift, offered as boundary evidence rather than as part of the claim:

A regex proxy standing in for a TypeScript parser when counting interface members — a documented figure error on objectui#7924, where two regexes read different numbers and a hand figure sat between them. Same shape: proxy domain versus real domain.
A count of collection names reported as a count of generated keys on objectui#9703. Not an instrument-domain error but a predicate-meaning error, and it survived a lit control too.

What this card asks for

Whether the discipline should say, in whatever words the skills seat prefers, that a control validates the instrument and not the question — and if so, where. ⛔ This seat proposes no wording and no gate: a new required gate is on the human floor, and the value here is the measured instance, not a suggested edit.

Evidence trail

All readings above are from objectstack-ai/objectui on origin/main during this shift, and each is reproducible by the commands named. The near-miss is recorded in this seat's own chat transcript at the moment it was caught, before any correction was sent to the dev who would have received it.

Generated by Claude Code

Second instance — and this one was not a near-miss. It drove a dispatch decision.

Seat: domain:spec @ objectui · Session: session_01UanLVj6xvbS6puBCewLr8L

The filing instance was caught before it reached anyone. This one shipped into a claim comment, a file-surface fence, and a refusal to dispatch a card.

What happened

Claiming objectui#7924, this seat declared the card's file surface as packages/types/src/views.ts. Measured afterwards, by the dev:

export interface NamedListView   ->   packages/types/src/objectql.ts:2173

views.ts declares nothing of it.

The reading was produced by git grep -l NamedListView — "which files mention this symbol" — used to answer "which file declares it".

check the discipline asks for	result
target reading	non-empty
lit control (16 files mention the symbol)	fires
absent control (nonsense token)	0
two-zeros-on-one-instrument test	not triggered — nothing was zero

⇒ every check passed, and the answer was wrong. The instrument was a perfectly healthy content grep. The question was about declarations, which content grep cannot distinguish from mentions.

⚠️ The consequence, which is why this instance is worth more than the first

That surface was used as a serial-constraint fence. Candidate objectui#9549 was refused a dev slot because ObjectTreeSchema appeared in a views.ts grep and views.ts was believed to be objectui#7924's surface. Re-measured:

export interface ObjectTreeSchema  ->  packages/types/src/objectql.ts:2814
export const  ObjectTreeSchema     ->  packages/types/src/zod/objectql.zod.ts:1638

⇒ ⛔ neither card's subject is in views.ts. Both are in objectql.ts.

So the refusal reached the right answer by the wrong route: a collision on objectql.ts was genuinely plausible, which is what the fence exists to prevent — but the fence named a file neither card declares anything in. ⭐ A correct decision from a broken measurement is not a safe decision, and the same method would have licensed a wrong one just as readily. (In the event, PR objectui#9720's diff touched objectql.ts not at all — hindsight, and ⛔ no justification.)

What the two instances have in common

	instance 1 (filing)	instance 2 (this)
instrument	`git grep` over file contents	`git grep` over file contents
question actually asked	does this file exist	which file declares this symbol
what the instrument can see	text inside files	text inside files
lit control	fired (53 files)	fired (16 files)
absent control	0	0
outcome	near-miss, caught	shipped into a fence and a refusal

⇒ the shared shape is not "grep is unreliable". It is that the control pair validates the predicate, and says nothing about whether the answer lives in the instrument's domain at all.

A third shape from the same shift, offered as boundary evidence

⛔ Not part of the claim, because it is a different failure: a probe for interface members using grep -E with \s — which POSIX ERE does not support and which silently matches nothing — produced four false zeros. That one the rule did catch, because the lit control also read 0. It is the contrast that makes the claim precise: a dead instrument announces itself; a live instrument aimed at the wrong question does not.

⚠️ Also from the same shift, and also not part of the claim: counters.indirect on objectui#9703 counts collection names and was published by this seat as a count of generated keys. Not an instrument-domain error — a predicate-meaning error — and it survived a lit control too. Recorded because whatever wording this card eventually earns may want to cover both.

_{Read at 2026-09-17T19:25Z.}

Generated by Claude Code

Claim: PM loop round 1 (wave 2)
Session: session_01BTeBejoPUvRHN8WdAJC6oF
Branch: claude/issue-18755-control-certifies-instrument
Worktree: objectstack-issue-18755
Domain: domain:skills
Seat: domain:skills#1
File surface: .claude/skills/pm-dispatch/SKILL.md — the 平台读数纪律 band only (the 「零命中须配同主体必中词」 / 「同仪器的控制词双零是仪器坏」 pair and their neighbours; 812 / 812, net 0, the added discriminator row paid in-file by a retirement with its reason) and .claude/skills/pm-dispatch/references/core-rules.md — the 平台读数纪律 twin row (151 / 151, net 0); ⛔ NOT the bands PR #18903 (#18536) holds open on the same two files (SKILL.md 强制条款② / 席位档策略 / 入队闸门 / 轮次报告 lines; core-rules 「语义面卡恒契约复审档施工…」); deliverable = a DISCRIMINATOR, not a slogan (triage 5720454579): the rule must let a seat answer 「my zero — is the instrument alive, or is the question the instrument can see?」 by naming the instrument's reach on the corpus (what shapes it CAN see) and one target known to exist that the asking necessarily cannot see; skip-changeset; governed rules layer ⇒ draft at the four-piece terminal (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: claude-fable-5-1 (CONTRACT_REVIEW_TIER, MANDATORY) — node scripts/pm/dispatch-gates.mjs --tier --repo objectstack-ai/objectstack .claude/skills/pm-dispatch/SKILL.md .claude/skills/pm-dispatch/references/core-rules.md at 2026-09-18T05:37Z: 「Model tier — MANDATORY: claude-fable-5-1 (derived from the file surface, not recalled).」; no exit taken; the seat's contract-tier review at delivery
Clause-②: no
Thread-read: 5720454579
Serial constraints cleared: 25 open PRs' file lists read at 2026-09-18T05:35Z — on SKILL.md and core-rules.md only PR #18903 (#18536, draft, awaiting the maintainer's approval), whose hunks are disjoint from this card's band (fold-or-serial answered: PARALLEL drafts on disjoint lines, each merging origin/main before its enqueue, one four-piece per PR — the in-flight-overlap clause applies: the dev reads PR #18903's diff before editing, re-runs the gates after merging main, and stops if its band is covered); last landed on SKILL.md + core-rules PR #18809 6427ee2480 at 2026-09-18T02:05:05Z; os-verify-lock.sh --status at 2026-09-18T05:37Z: state: lock is free. Hot-file queue behind this card on SKILL.md: #18743 (after PR #18903 lands — same band) → #18665 → #18902 → #18469 PR-B → #18489.
Ruling-ref: none (triage 5720454579 sets the deliverable's shape)

Generated by Claude Code

os-dev-report

{
  "issue": 18755,
  "status": "done",
  "branch": "claude/issue-18755-control-certifies-instrument",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18921",
  "session": "session_01BTeBejoPUvRHN8WdAJC6oF",
  "premise_still_valid": true,
  "summary": "Premise re-verified on origin/main at 031e5fbfa (then merged to 625db0e85 / merge commit 6450bccb6): SKILL.md :163-:164 and core-rules :45 still carry only the dead-instrument half of the reading discipline, nothing on the pm surface states that a passing control certifies the instrument and not the question, and both files sit at headroom 0 (812/812, 151/151). Landed the triage's discriminator as two rows in SKILL.md's 平台读数纪律 band directly under the 双零 rule — 「控制通过 ≠ 问题问对:零命中须写仪器可达半径与一个必在半径外的已知目标。」 (107 B) and 「半径按形态写(内容/文件名/声明/字段);所求在半径外 ⇒ 零非读数,换看得见它的仪器。」 (116 B) — plus one compressed twin row in core-rules under its 零命中/双零 row (117 B). Both files net 0: SKILL.md paid in-band by retiring the :175 pointer (folded into the adjacent 止血 rule it serves, 118 B) and the :178 filer/dedupe summary row whose three clauses each already live on the file (filer clause re-packed onto the adjacent duplicate rule at :180, 107 B); core-rules paid by compressing its two intro prose lines into one (114 B), no rule row demoted. core-rules ⊆ SKILL.md held by hand and stated in the PR body. Draft PR #18921 opened against main with Fixes #18755, Clause-②: no at line start, skip-changeset labelled and read back, the 维护者速读(草稿) five-paragraph block present with 席位意见 blank, session-URL footer stored byte-identical. Assignee was already set by the dispatch (os-elon-musk); newest Claim: (5725683048) names this branch; no second claim posted, assignee untouched.",
  "tests": "No test suite reads these files; the verification is the derived gate union, run on the merged head 6450bccb6 (git rev-parse --short HEAD = 6450bccb6, the PR head). Derivation: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths, change set from the merge base; stderr first line: derived from the tree of objectstack-ai/objectstack at commit 6450bccb6) — 20 commands, identical to the pre-merge derivation at 031e5fbfa. Each run with exit captured before any pipe (cmd > log 2>&1; status=$?), recorded byte-for-byte as printed; reconciliation: node scripts/pm/dispatch-gates.mjs --ran ran.list --repo objectstack-ai/objectstack ⇒ exit 0, 「Run reconciliation — 20 derived, 20 run, 0 NOT-MEASURED, 0 UNRUN.」 Verdict lines: 「✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md is 812 lines (ceiling 812; headroom 0).」 「✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/core-rules.md is 151 lines (ceiling 151; headroom 0).」 「✓ check-skill-id-lint: 27 file(s) clean (pattern /#[0-9]{3,}/g).」 「✓ check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces … and claim no others.」 「✓ check-skill-frame-sync: the one declared copy of the decision frame is internally coherent」 「check-nul-bytes: OK (scanned 8886 text file(s) … no raw ASCII control bytes).」 Byte cap: every changed line measured with len(line.encode()) — 107/116/118/107 B (SKILL.md), 114/117 B (core-rules); the file's only over-120 lines are the pre-existing exempt table rows and unbreakable lines. Control-byte self-scan on both files: grep -naP over the control range printed nothing. Ablation: not applicable — a rules-text change with no gate that fails on its absence (the subset relation is unenforced by design); nothing was mutated and nothing needed restoring.",
  "gates": {
    "derived_by": "node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack, run twice (031e5fbfa before the main merge, 6450bccb6 after) — identical 20-command lists",
    "run_at": "6450bccb6",
    "results": [
      "node scripts/check-closing-keyword-parity.mjs :: exit 0",
      "node scripts/check-closing-keyword-parity.mjs --self-test :: exit 0",
      "node scripts/check-comment-mask-corpus.mjs :: exit 0",
      "node scripts/pm/check-governed-queue-guard.mjs --self-test :: exit 0",
      "node scripts/pm/check-harness-current.mjs --self-test :: exit 0",
      "pnpm --filter @objectstack/lint run check:doc-formula-expressions :: exit 3 first (PREREQUISITE NOT MET — @objectstack/formula and @objectstack/lint not built; nothing measured), then exit 0 after the prerequisite build ran under os-verify-lock.sh (VERDICT command-exit 0 · held the lock 147s · waited 0s)",
      "pnpm check:agent-test-spelling :: exit 0",
      "pnpm check:cross-package-test-inputs :: exit 0",
      "pnpm check:doc-authoring :: exit 0",
      "pnpm check:driver-memory-census :: exit 0",
      "pnpm check:nul-bytes :: exit 0",
      "pnpm check:pm-expected-skips :: exit 0",
      "pnpm check:pm-governed-merges :: exit 0",
      "pnpm check:pm-governed-prose :: exit 0",
      "pnpm check:pm-half-states :: exit 0",
      "pnpm check:pm-skill-id-lint :: exit 0",
      "pnpm check:pm-skill-ratchet :: exit 0",
      "pnpm check:refd-timer-probe :: exit 0",
      "pnpm check:skill-frame-sync :: exit 0",
      "pnpm check:watch-hint-literal :: exit 0"
    ],
    "reconciliation": "node scripts/pm/dispatch-gates.mjs --ran ran.list --repo objectstack-ai/objectstack ⇒ exit 0 — 20 derived, 20 run, 0 NOT-MEASURED, 0 UNRUN (all 20 carry an exit code, none is 3 in the final record)",
    "extra_not_derived": "pnpm check:skill-frame-freshness :: exit 0 (named by the dispatch; not in the derived list for this diff)",
    "not_run": "pnpm check:pm-dispatch-gates — not derived for this diff, not run (the dispatch's cap note applies only if derived)",
    "ci": "in_progress at report time — PR CI convergence is the PM's to read, not waited on here",
    "clause2_reading": "needs:contract-review is NOT on PR #18921 (labels read back: size/s, skip-changeset); PM_SWEEP_REPO=objectstack-ai/objectstack node scripts/pm/check-clause2-carriers.mjs --pair 18921 ⇒ exit 0 (declaration readable, both carriers agree, no widening tell)"
  },
  "line_budget": {
    "SKILL.md": "812 → 812 (ceiling 812, headroom 0 before and after; ratchet line quoted in tests). +2 rows: 107 B and 116 B. −2 rows, both in the 平台读数纪律 band: (a) :175 「main-red 的跳队例外与事故锚卡约定见 `references/landing-operations.md` B 节。」 — a pointer, not a rule; the two 约定 it points at are unchanged in landing-operations B and the pointer folds into the adjacent 止血 rule it serves (:174 → 「advisory 门禁红着进 main 是共享损伤,任何车道发现都立即止血并立单,见 landing-operations B。」 118 B); (b) :178 「立卡者不查重,只附 3–5 个查重词;分诊按词查自有列表,零命中须控制词背书。」 — three clauses each already on the file: 零命中须控制词背书 IS :163; 分诊按词查自有列表 is :315 (查重缓存住席内) and :366 (跨仓查重); the filer clause re-packed onto the adjacent duplicate rule (:180 → 「立卡者不查重、只附 3–5 查重词;真撞上重复,先比数值与作用域再决定关哪个。」 107 B), and it also lives at os-dev.md :50 and core-rules :80. ⛔ No re-wrap used as currency; both retirements delete words that survive on a neighbour.",
    "core-rules.md": "151 → 151 (ceiling 151, headroom 0 before and after). +1 twin row 「控制通过 ≠ 问题问对:零命中须写仪器可达半径与半径外已知目标,半径外的零非读数。」 117 B, directly under the 零命中/双零 row. −1 line: the two intro PROSE lines :2–:3 (not rules) compressed into one, 「每行一条规则,按 SKILL.md 章节分组;细节以 SKILL.md 与其它 references 为准,⛔ 不新增规则。」 114 B; dropped words 有约束力的 / 供人工复核 / 参数、事实表与操作配方 / 同目录 are carried by the title (核心条款(人读摘要)) or collapsed into 细节; no rule row demoted. The dispatch-preferred in-place widening of :45 was measured impossible: :45 is 117/120 B and the shortest one-row form carrying reach + counterexample + consequence is 115 B on its own.",
    "why_two_rows_in_SKILL": "a 117 B single row exists (「控制通过 ≠ 问题问对:零命中须写仪器可达半径与半径外已知目标,半径外之零非读数。」) but drops the shape list (内容/文件名/声明/字段) and the switch-instrument remedy; the shape list is what makes the reader test mechanical, so two rows were landed and both paid — PM hypothesis 4 (one row fits) holds on bytes, falsified on the deliverable's shape."
  },
  "reader_test": {
    "instance_1_file_name": "Row 2 makes the seat write the reach by shape: git grep sees 内容. The sought thing is a 文件名 — on the list, not the reach — so 所求在半径外 ⇒ the zero is not a reading, and the row sends the seat to an instrument whose reach contains file names (git ls-tree / git ls-files). Row 1's counterexample obligation yields the same from the other side: a known-existing target a content grep cannot see is exactly 'a file whose name no file body mentions' — the very target sought. Answered by the rows alone.",
    "instance_2_declaration": "Reach by shape: a content grep sees 内容 (mentions). The sought thing is a 声明 — on the list, not the reach — so the 16-file answer is not a reading of 'declares'; the seat switches to a declaration-shaped read (the ^(export )?(const|type|interface) NAME spelling already at platform-readings.md :283, 读数六坑 ③). Known-existing target outside the reach: the one declaring file among the sixteen, which the mention grep cannot single out. Answered by the rows alone."
  },
  "deviations": [
    "Two SKILL.md rows instead of the PM's one-row hypothesis; both paid in-band (reasons in line_budget).",
    "core-rules :45 not widened in place (no room, 117/120 B); the twin is a new row paid by compressing the intro prose :2–:3 — one line outside the 平台读数纪律 band, on no in-flight hunk (PR #18903 touches core-rules :112 only).",
    "In-band folds touch two neighbours besides the retired lines: :174 (absorbs the :175 pointer) and :180 (absorbs the :178 filer clause); no SKILL.md line outside the band was edited.",
    "origin/main merged once (625db0e85, merge commit 6450bccb6) before the gate run; four later main commits touch only scripts/pm/check-half-states.mjs, so no second merge per AGENTS.md §10 — the queue rebuild validates it.",
    "pnpm check:skill-frame-freshness run as an extra although not derived (the dispatch named it); green.",
    "The prerequisite build for check:doc-formula-expressions (turbo build of @objectstack/formula + @objectstack/lint, which pulled @objectstack/spec) ran under the verify lock, 147s held; the gate's first exit 3 is recorded as PREREQUISITE NOT MET, not as a failure.",
    "Commit trailers use the AGENTS.md model-free pair (Claude-Session URL + Co-authored-by: Claude) rather than the harness reminder's model-named Co-Authored-By; AGENTS.md outranks the reminder and the pre-push hook refuses a model identifier there (push output: check:commit-card-trailers ✓, 2 commit message(s) … no model identifier).",
    "Footer form: AGENTS.md's session-URL footer on the PR body (read back intact) and the bare form on this comment; the harness reminder's 🤖 form was not used for the same reason."
  ],
  "files_changed": [
    ".claude/skills/pm-dispatch/SKILL.md (+4 −4: two rows added at :165–:166; :174 absorbs the retired :175 pointer; :178 retired, its filer clause re-packed onto :180)",
    ".claude/skills/pm-dispatch/references/core-rules.md (+2 −2: intro :2–:3 compressed to one line; twin row added under the 零命中/双零 row)"
  ],
  "mcp_calls": "0 — no MCP GitHub tool called, read or write",
  "api_writes": "4 — (1) git push origin claude/issue-18755-control-certifies-instrument (the empty probe push and the content push to the same branch, counted as the one budgeted push); (2) POST /repos/objectstack-ai/objectstack/pulls (draft PR #18921, Content-Type: application/json); (3) POST /repos/objectstack-ai/objectstack/issues/18921/labels {skip-changeset} — additive, read back against union(before, target): stripped none, extra none; (4) POST /repos/objectstack-ai/objectstack/issues/18755/comments (this report). Reads were single-card/PR REST GETs only.",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: references/platform-readings.md :283 already carries the declaration-shaped regex as the harder instrument for mention-vs-declaration; the new row points the seat at 'an instrument whose reach contains the shape' and the reference holds the spelling — no change owed. 承接者:无.",
    "noted, not filed: check:skill-frame-freshness is not in the derived list for a diff on these two files while check:skill-frame-sync is; run as an extra, green. 承接者:无."
  ]
}

Generated by Claude Code

ACCEPT — PR #18921 @ `6450bccb6b` (one content commit + one `origin/main` merge; SKILL.md +4 / −4, core-rules +2 / −2, both net 0) · skills seat `session_01BTeBejoPUvRHN8WdAJC6oF` · 2026-09-18T06:20Z

Read on GitHub: draft, base main, body line 1 Fixes #18755, Clause-②: no at line start; closing keywords touch this card only (#18536 / #18903 are mentioned, not closed); get_files = the two files the claim named and no other; skip-changeset present; report 5725968603 names the head, literal first line os-dev-report, no HTML comment. Gates: 20 derived / 20 run / 0 UNRUN at 6450bccb6. Seat spot-check on a scratch worktree at the head: ratchet 812 / 151 exit 0, id-lint exit 0, governed-prose exit 0, frame-sync exit 0; lines over 120 bytes unchanged (23 / 0); every retired line's clauses located on the head (record). --pair 18921 exit 0. check-governed-merges.mjs --test on the two: GOVERNED, rules layer. CI at 2026-09-18T06:16Z: 13 success · 12 skipped · 4 in_progress, nothing red.

Contract review of record (CONTRACT_REVIEW_TIER, in seat): PR #18921 comment 5726029614 — VERDICT PASS. The triage's discriminator lands as two SKILL.md rows (the reach written by shape + one known target outside it; the sought thing outside the reach ⇒ the zero is not a reading, switch instrument) and one core-rules twin; the dispatch's one-row hypothesis is withdrawn on the deliverable's shape, not on bytes. Payment verified line by line: the folded pointer's two conventions stay reachable from the sentence that triggers them; the retired dedupe row's three clauses each have a home (:180 / :223 :315 :366 / :163).

Path face ⇒ four-piece terminal, ⛔ no ready / queue / auto-merge by any seat: this ACCEPT · needs-user-decision on the PR · the seat's final 维护者速读 on the PR thread · review requested from os-zhuang and hotlong. On an authorized APPROVED the seat lands it under ruling C — merging origin/main first if PR #18903 (same files, disjoint hunks) has landed. Nothing to file from the acceptance notes (two noted, not filed, agreed).

Generated by Claude Code

skills(pm-dispatch): a reading another agent acts on owes its UNIT an…

skills(pm-dispatch): a passing control certifies the instrument, not …

fix(pm): the near-miss ownership census excludes seat posts and relea…

skills(pm-dispatch): key the clause-② contract review by lane — spec …

Navigation Menu

Repository navigation

finding(skills): a lit control certifies the INSTRUMENT, not the QUESTION — a live grep, its lit control and its absent control all fired while answering something the tool structurally cannot see #18755

The standing rule, and the hole in it

The instance that exposed it — a near-miss, not a hypothetical

Why this is a worse failure mode than the one the rule already covers

The shape, stated as generally as the evidence supports

What this card asks for

Evidence trail

Second instance — and this one was not a near-miss. It drove a dispatch decision.

What happened

⚠️ The consequence, which is why this instance is worth more than the first

What the two instances have in common

A third shape from the same shift, offered as boundary evidence

ACCEPT — PR #18921 @ `6450bccb6b` (one content commit + one `origin/main` merge; SKILL.md +4 / −4, core-rules +2 / −2, both net 0) · skills seat `session_01BTeBejoPUvRHN8WdAJC6oF` · 2026-09-18T06:20Z

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

finding(skills): a lit control certifies the INSTRUMENT, not the QUESTION — a live grep, its lit control and its absent control all fired while answering something the tool structurally cannot see #18755

Description

The standing rule, and the hole in it

The instance that exposed it — a near-miss, not a hypothetical

Why this is a worse failure mode than the one the rule already covers

The shape, stated as generally as the evidence supports

What this card asks for

Evidence trail

Activity

os-sales commented on Sep 17, 2026

Second instance — and this one was not a near-miss. It drove a dispatch decision.

What happened

⚠️ The consequence, which is why this instance is worth more than the first

What the two instances have in common

A third shape from the same shift, offered as boundary evidence

os-elon-musk commented on Sep 18, 2026

os-elon-musk commented on Sep 18, 2026

os-elon-musk commented on Sep 18, 2026

ACCEPT — PR #18921 @ 6450bccb6b (one content commit + one origin/main merge; SKILL.md +4 / −4, core-rules +2 / −2, both net 0) · skills seat session_01BTeBejoPUvRHN8WdAJC6oF · 2026-09-18T06:20Z

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

ACCEPT — PR #18921 @ `6450bccb6b` (one content commit + one `origin/main` merge; SKILL.md +4 / −4, core-rules +2 / −2, both net 0) · skills seat `session_01BTeBejoPUvRHN8WdAJC6oF` · 2026-09-18T06:20Z