ums: hold a fallback self-review to the bot's prose-fact-check bar

[dem-ucdh]

Closes #1093.

What this records

Sharpens CLAUDE.md's "Do the review yourself when the @claude workflow doesn't produce a verdict" section with a Do/Don't-labelled note about a specific trap: a fallback self-review -- one posted because the automated reviewer was unavailable (quota-skipped, stub, or erroring on an infra failure) -- feels like a stopgap and so tends to get a shallower pass than the round deserves. The gap is predictable: a shallow self-review checks structure (dogfood back-reference, ASCII, semantic line breaks) and skips the prose fact-check, so a false mechanism claim or a misattributed citation sails straight through.

The fix: run the applicable prose-review skills against the diff's own factual claims -- fact-check-prose, the cause claim-type check in metacognitive-monitoring, and the read-the-cited-source rule in address-every-comment -- not just the structural checks. The standard does not relax because the reviewer it replaces was absent.

Case record

#1092, 2026-08-03: a fallback self-review posted while claude-review was erroring on infra failures reported "no findings / ready" and missed two content bugs the recovered bot then caught -- a false mechanism claim about a directional-word grep, and a citation misattributed to shared/writing/definition-crossrefs.md (which is about formal Quarto crossref-div ordering and says no such thing). Both are semantic errors escaping mechanical checks -- the very subject of the reviewed section.

Placement / near-duplicate check

• CLAUDE.md "Do the review yourself..." is the closest home; it already says "Apply the same review standards the bot would" but does not warn about fallback shallowness or name the prose fact-check skills. This is the fallback-specific sharpening of that line.
• shared/workflow/ardi.md's pre-push self-review section names fact-check-prose, but for the general pre-push convention check, not the fallback-substitute case or its stopgap-shallowness trap.
• shared/workflow/fully-clean.md covers self-review-as-fallback but framed around "still need an external verdict when reachable", not the prose-fact-check bar for the fallback you post.
• shared/workflow/metacognitive-monitoring.md supplies the cause claim-type check the fix points at (its claim-type list is State/Scope/Cause/unexamined-default); the citation-checking rule lives in shared/workflow/address-every-comment.md, so the fix links both.
• Normalized searches for stopgap, shallow, same bar as the bot, same review standards found no existing coverage of this trap.

Collision note

Disjoint from the other open PRs. #1088 also touches CLAUDE.md but at a distant section (~L521, the review-verdict-query section); this edit is at ~L887. The other open PRs (#1047 hooks/, #1091/#1089/#1087 shared/, #1092 shared/writing/forward-references.md, #1042 hooks/) do not touch CLAUDE.md.

Self-checks

• Added lines are pure ASCII (no em/en-dash, curly quotes, or U+00D7); --- used for dashes.
• One sentence per line; the **Bold.** + next-sentence pattern matches the dominant corpus convention (the local reformatter's suggestion to merge them is the documented .**-boundary false-merge it makes on all 31 existing bold-lead entries too).
• All referenced fragment paths verified to exist; the definition-crossrefs.md characterization in the case record was verified by reading the file.

Review round 1 (addressed)

claude-review correctly caught that the paragraph itself misattributed a "citation" claim-type check to metacognitive-monitoring.md (which has no such type) -- the exact bug the paragraph warns about. Fixed in 45fdb5c: the file's main sentence and first Do bullet now point the cited-source half at address-every-comment, and this PR body is synced to match.

[claude]

Code review

Reviewed the diff (one new subsection added to CLAUDE.md, no code changes). Verified: both linked files exist (shared/writing/fact-check-prose.md, shared/workflow/metacognitive-monitoring.md), the cited case record (#1092, #1093) accurately reflects that PR's history, definition-crossrefs.md is correctly characterized, the Do/Don't block is properly labelled per CLAUDE.md's "Record both the pattern and the anti-pattern" convention, added lines are ASCII-only with no banned punctuation, and the insertion doesn't break any downstream count-based back-reference or duplicate the nearby "Do the review yourself..." / "Self-review is the immediate fallback..." material.

One finding, posted inline on CLAUDE.md:890: the new text credits shared/workflow/metacognitive-monitoring.md with a "citation" claim-type check, but that file defines exactly four claim types (State, Scope, Cause, An unexamined default) and contains zero occurrences of "citation." The citation-verification rule actually lives in shared/workflow/address-every-comment.md (and is touched on in shared/writing/fact-check-prose.md, already linked one line up). This is the same misattributed-citation defect the paragraph itself is warning readers to catch — worth a quick fix before merge.

Verdict

Needs minor changes — one non-blocking-but-ironic citation fix (inline comment above); otherwise ready for merge.

[github-actions]

Claude finished review — View run

---

Review posted. Summary:

Finding (1, inline): CLAUDE.md:890 credits shared/workflow/metacognitive-monitoring.md with a "citation" claim-type check, but that file only defines four claim types — State, Scope, Cause, and "An unexamined default" — with zero mentions of "citation" (confirmed via grep and independently by three separate review agents). The citation-verification guidance actually lives in shared/workflow/address-every-comment.md. This is notably the exact class of bug — a misattributed citation — that the new paragraph itself is warning readers to watch for.

Everything else checked out clean: both linked files exist, the cited case record (#1092/#1093) accurately reflects history, the Do/Don't block is properly labelled, punctuation/line-break conventions are followed, and the insertion doesn't duplicate nearby content or break any back-references.

Verdict

Needs minor changes — one non-blocking citation fix flagged inline (with a suggested rewording); otherwise ready for merge.

[github-actions]

💰 Cost: $13.9774 (review) — run

[github-actions]

Claude finished review — View run

---

Code review — #1094

Scope checked: --comment was not passed, so per the code-review skill's step 7 I stopped short of posting anything to GitHub and am reporting findings here only.

What changed since the last round: The PR added a new CLAUDE.md subsection ("A fallback self-review is prone to being shallow...") plus a Do/Don't block and a case record citing #1092/#1093. My prior review (posted 2026-08-03T10:42:31Z, comment) found one issue: the paragraph credited shared/workflow/metacognitive-monitoring.md with a "citation" claim-type check it doesn't have.

Verification of the fix: Commit 45fdb5c re-split the sentence so metacognitive-monitoring is now cited only for the cause claim-type check, and the "read the cited source" rule is now attributed to shared/workflow/address-every-comment.md. I independently re-verified:

• metacognitive-monitoring.md defines exactly State/Scope/Cause/An-unexamined-default, 0 occurrences of "citation" (grep -ic citation → 0).
• address-every-comment.md:863 does contain "When a finding cites a source, read the cited source before reproducing anything" — the citation now resolves correctly.
• The parallel first Do bullet was fixed the same way.
• The "above" back-reference to "Apply the same review standards the bot would" correctly points to line 859, which is above the insertion in the same section (no forward-reference / stale-back-reference issue, per the very rule class this PR is about).
• The case record's claims about ums: inserting prose can stale a downstream count back-reference #1092 and A fallback self-review is prone to being shallow; hold it to the same prose-fact-check bar as the bot it replaces #1093 match those items' actual bodies.
• Added lines are ASCII-only (no em/en-dash, curly quotes, ×) and one-sentence-per-line, consistent with the rest of the file.
• No duplication with the adjacent "Self-review is the immediate fallback..." material — that block is about whether an external verdict is still required; this new block is about the depth of the self-review itself.

No new issues found in this round.

Verdict

Ready for merge.

[github-actions]

💰 Cost: $5.2835 (review) — run