Published-skills factual sweep: verify every behavioral claim in skills/** against the implementation — program anchor

[zhuangjianguo]

Program anchor, filed by the skills lane seat (session session_01EXxTW8mvPBhoHxmyPZ63de) on the maintainer's scheduling order.

Mandate (verbatim)

2026-08-31, chat: 「指对外发布的 skills/**，也需要排程」 — confirming the program proposed after PR #13577. Standing context, 2026-08-21: 「对外发布的 skills 是整个平台的最大价值」.

Why a sweep, not more incidents

The published corpus has token/line ratchets (scripts/check-skills-token-ratchet.mjs) and compile-validity gates (check:skill-examples type-checks 260 prose examples), but nothing verifies behavioral truth against the implementation. PR #13577 proved the class: two published rows taught $exists → MongoDB $exists while every engine implements has-a-value (!= null) — found incidentally by an unrelated card, not by any sweep. One measured false row in ~186k published tokens is a floor, not a ceiling.

Method — PR #13577 is the spec

Per behavioral claim (operator/API table rows, key names, behavioral sentences, code examples' asserted outputs):

1. Locate the implementing code (never verify one document against another document);
2. verify by reading plus an executed test where the claim is behavior-bearing;
3. verdict per claim: VERIFIED / FALSE (fixed in the sweep PR itself, byte-neutral-or-shrinking under the token ratchet — ⛔ no ceiling raises by devs; nuance overflow goes to content/docs/** follow-ups, not into ratcheted skill text) / NOT MEASURABLE (recorded with the reason — never silently skipped);
4. non-vacuity control per sweep: at least one claim proven true by execution — a sweep reporting only-verified with zero executed evidence has measured nothing;
5. PR body carries the per-item 落点 | before | after list (sweep-packaging五条); zero out-of-list changes.

Inventory and order

#	package	.md lines	note
①	objectstack-formula	579	calibration flight — measures cost-per-claim and false-density to size the rest
②	objectstack-data	4,935	largest; same semantic family as the proven false row; split-by-file allowed with report
③	objectstack-query	1,550	⛔ serial behind PR #13577's merge (same files)
④	objectstack-platform	2,633
⑤	objectstack-ui	2,298
⑥	objectstack-automation	1,163
⑦–⑪	api 707 · i18n 807 · upgrade 696 · ai 691 · pm-dispatch 1,077
⑫	objectui skills/objectui	5,686	merge regime pending the ui#6866 half-1 ruling; until ruled, governed posture (the fleet-side default)

Batching: one flight per package, ≤3 parallel, no file overlap. All objectstack skills/** PRs: governed — draft, human merge.

Members are filed one per package at dispatch time and link back here; this anchor tracks the roster. First calibration results re-size the schedule and get posted here.

Refs: PR #13577 (the class proof and the method precedent) · #13539 (the incident card) · #13597 (the separate PM-corpus audit — different corpus, parallel program).

Generated by Claude Code

[zhuangjianguo]

Roster update: flight ⑫'s merge regime is RESOLVED, not pending — ui#6866 half-1 was already ruled on 2026-08-30 (决裁批 #7, maintainer verbatim 「批 #7 同意」, recorded at ui#6866 comment 5469339478): Option A — published skills/** is governed in objectui too; .md content ⇒ maintainer review, pure code faces ⇒ seat contract-tier self-review. My earlier "pending" note came from a body-only triage read; corrected at the source card (comment 5474282135). Flight ⑫ therefore runs under the same governed-draft posture as flights ①–⑪. The half-1 TEXT implementation in objectui AGENTS.md is a separate card on that ruling, serial behind ui PR #6976.

---

Generated by Claude Code

[zhuangjianguo]

Calibration results (flight ① delivered — PR #13674, ACCEPTed, awaiting human merge)

• 197 behavioral claims across 13 classes in 579 lines (~118 lines non-behavioral, excluded by rule) · 6 FALSE = 3.0% · 26 NOT MEASURABLE (recorded) · 165 VERIFIED, all with executed probes where behavior-bearing.
• The sizing finding: falsehoods concentrate ~8× in surfaces/mapping tables — 4/38 (10.5%) there vs 2/159 (1.3%) elsewhere. Matches PR docs(skills): $exists means HAS A VALUE — retire the MongoDB $exists claim from objectstack-query #13577's pattern (a mapping column asserting engine behavior). Flights ②–⑫ therefore verify TABLES with execution priority first, prose classes second; the method stays full-inventory (the 1.3% is still real falsehoods).
• Falsehood flavors worth pattern-matching in later flights: retired-surface rows outliving ADR-0049 removals (2 of 6) · phantom fields with zero repo-wide hits (2) · slot-semantics overclaims (defaultValue not auto-wrapping) (1) · unshipped-path pointers (src/ in a dist-only package) (1).

Revised schedule: mid-size packages (≤1,200 lines) = one flight each; objectstack-data (4,935) = one card, up to three sequential PRs at whole-file boundaries (Part-of until the last, which closes); first full pass still projected 2–4 days, gated on merge cadence. Flight ② dispatching now.

---

Generated by Claude Code

[zhuangjianguo]

Calibration addendum from flight ①'s full report (effort economics, for sizing ②–⑫):

claim class	effort/claim	note
executable-oracle rows (stdlib, formatters, cron forms, translation tables)	~1.5 min	they BATCH: one probe file, one run
spec-surface rows	~4 min	schema location + retirement-note chasing — where both the cost and the yield live
prose behavioral paragraphs	~8 min	truth lives in large lint/validate modules
cross-package runtime semantics	NOT MEASURABLE from a package flight	needs a kernel harness; ~20 min/claim if a flight is scoped to build one — 26 such claims recorded in flight ① and this class will recur in data/platform; a dedicated harness flight may pay for itself if their count crosses ~50 program-wide

Wall clock flight ①: ~3.5h incl. builds (~45 sec/claim averaged — misleading, use the class split). A table-first pass would have found 4 of the 6 falsehoods for ~20% of the effort — ②–⑫ briefs already order it so.

Mechanical follow-on carded: #13679 — a named-identifier liveness gate over skills/** (3 of 6 falsehoods were phantom rows, each detectable by one grep). The sweep fixes today's corpus; the gate keeps tomorrow's ADR-0049 retirements from re-creating phantoms.

---

Generated by Claude Code

[zhuangjianguo]

Correction to the addendum above: the named-identifier liveness gate card is #13678 (the comment guessed #13679 before the create returned — read all references to #13679 there as #13678).

---

Generated by Claude Code