Verify Claims

/verify-claims — Chain-of-Verification on a Draft

Fact-check a draft using the Post-Flight Verification protocol (.claude/rules/post-flight-verification.md).

Input: $ARGUMENTS — path to a file containing the draft (markdown, .qmd, .tex, .md) or a shorthand pointer. Optional flags:

• --source <path-or-url> — one or more source-material pointers (repeat for multiple). If omitted, the skill infers from context (e.g., papers referenced, cited arXiv URLs).
• --no-fail-closed — downgrade FAIL outcomes to warnings without regeneration. Use sparingly.

When to pick this skill

• /verify-claims (this skill) — ad-hoc fact-checking on any draft or text block the user hands you. One-shot, user-invoked.
• Other skills that auto-run Post-Flight internally (/lit-review, /research-ideation, /respond-to-referees, /review-paper --peer) — no need to call this separately; they already run it.
• /proofread — grammar, typos, overflow. Different lens.
• /review-paper (default mode) — full manuscript review, not just claim verification.
• /validate-bib — checks citations *exist* and are well-formed (structural + DOI). This skill checks they *hold* (the cited paper supports the attributed claim). Complementary — run both before submission.

How it works

Implements the 4-step CoVe loop from Dhuliawala et al. 2023 (arXiv:2309.11495), with architectural enforcement of the fresh-context independence trick.

Phase 0 — Pre-Flight

Confirm:

• Draft file exists and is readable
• At least one source pointer available (either --source or auto-detected from draft)
• claim-verifier agent file exists at .claude/agents/claim-verifier.md

If any fail → surface the failure, do NOT proceed.

Phase 1 — Extract claims

Read the draft. Identify factual assertions of these types:

| Type | Example |

|------|---------|

| Citation | "Smith (2019, *JEL*) shows X" |

| Numerical fact | "N = 10,000", "ATT = 0.42" |

| Negative literature | "No prior work studies X" |

| Named entity | researcher, paper title, venue, package, estimator name |

| Dataset claim | "The CPS contains field educ_attain" |

Skip: opinions, forward-looking suggestions, definitions the draft introduces.

For citation-type claims, extract the claim↔citation PAIR — not just the citation. Capture *what* the draft attributes to *which* work, so the verifier checks appropriateness (does Smith 2019 actually *show* X?), not merely existence. "Smith (2019) shows a positive wage effect" becomes {cite: Smith2019, attributed: "positive wage effect"}. This is the layer /validate-bib explicitly defers here: validate-bib confirms the citation *exists and is well-formed*; this skill confirms it *holds*. A mis-citation (the paper exists but says something else, or the opposite) is exactly a numeric/directional contradiction → HIGH-WARN unless a concrete author_alternative is recorded (then EXPLAINED).

Output a claims table:

| ID | Claim | Source hint |
|----|-------|-------------|
| C1 | ... | ... |

Phase 2 — Generate verification questions

One question per claim. Make it specific and answerable from the source alone.

Phase 3 — Spawn claim-verifier (forked, fresh context)

Task: subagent_type=claim-verifier, context=fork
Prompt: hand over claims table + verification questions + source material pointers.
        Do NOT include the draft text.

The forked agent runs the CoVe independent-answer step. It has never seen the draft and cannot confirm-bias. It returns a structured verification report.

Phase 4 — Reconcile

The verifier returns a per-claim verdict in one of these severity tiers:

• HIGH-WARN — fabricated reference (the cited paper doesn't exist at the named venue/year), draft claim directly contradicted by the source, or not_found retrieval that the verifier interprets as a hallucinated citation. Gate-refuse — these block /commit for any file /verify-claims was just run against, unless the user explicitly overrides with --no-fail-closed or sets verifyClaims.allowHighWarn: true in .claude/settings.json.
• MED-WARN — transient infrastructure / retrieval failure (paywall the verifier can normally bypass via cached metadata; DOI resolver timeout; partial PDF read). Surface for the author; do not gate-refuse.
• LOW-WARN — source genuinely inaccessible (paywalled and not in cache; private dataset; pre-print server transient). Surface with cannot-verify flag; do not gate-refuse.
• EXPLAINED (v2.0) — a numeric/directional contradiction the author has *pre-justified* with a concrete named alternative (different defensible edition, specification, sample, or rounding convention), passed to the verifier via the claim's author_alternative field. Surfaced with the evidence and the recorded reason; non-gating. The hard floor holds: a *fabricated* citation is never EXPLAINED, and a blank/vague alternative stays HIGH-WARN. This mirrors audit-reproducibility's EXPLAINED disposition for numeric claims — a mismatch is not always a failure when a defensible alternative is named.

Verdict aggregation by tier across all extracted claims (EXPLAINED counts as non-gating, like LOW):

| Tier counts | Outcome | /commit behaviour |

|---|---|---|

| 0 HIGH, 0 MED, ≥ 0 LOW/EXPLAINED | PASS (green block) | proceeds |

| 0 HIGH, ≥ 1 MED, any LOW/EXPLAINED | PARTIAL (yellow block) | proceeds with warning |

| ≥ 1 HIGH | FAIL (red block) | halts unless override |

--no-fail-closed opts out of the gate-refuse behaviour on HIGH-WARN. Use sparingly — it's there for offline / hallucination-sensitive contexts where the user accepts the risk in writing.

If the draft is writeable and the user asked for auto-correction, regenerate the affected sections using the verifier's evidence. Otherwise return the report and let the user decide.

Example

/verify-claims quality_reports/lit-review_staggered-did.md --source master_supporting_docs/callaway_santanna_2021.pdf --source master_supporting_docs/dechaisemartin_dhaultfoeuille_2020.pdf

Expected output (abridged):

## Post-Flight Verification — lit-review_staggered-did.md

**Claims extracted:** 14
**Verified independently:** 14 (forked claim-verifier)
**Outcome:** PARTIAL — 12 verified, 1 discrepancy, 1 unverifiable

### Discrepancies

- **C7** — draft claims "de Chaisemartin & D'Haultfœuille (2020) *propose* a DR estimator." Source Section 4 shows they propose a weighting estimator, not DR. Recommend correction.

### Unverifiable

- **C12** — draft cites "Borusyak et al. 2024 (working paper)". No canonical URL in provided sources. Recommend user supply DOI or arXiv link.

### Verified

| ID | Claim | Evidence |
|----|-------|----------|
| C1 | "Callaway & Sant'Anna 2021 use group-time ATT" | p. 5, eq. (3) |
| ... | ... | ... |

Fail modes and recovery

Verifier times out: surface a warning block, return draft as provisional. Do not silently ship.

Source material inaccessible (paywall, 404): report the specific claims that hinge on it, flag as cannot-verify, recommend user supply an alternative source.

Draft contains only opinions / forward-looking text: report "no verifiable factual claims extracted — nothing to check" and return.

Cross-references

• .claude/agents/claim-verifier.md — the forked verifier.
• .claude/rules/post-flight-verification.md — the protocol.
• MEMORY.md [LEARN:pattern] on Chain-of-Verification vs critic-fixer vs cross-artifact review.