| name | paper-claims |
| description | Validate, extract, and verify manuscript claims against evidence. Use when editing .tex files, before submission, or when asked about claims/evidence/provenance. |
| allowed-tools | Read Grep Glob Bash Agent |
| effort | high |
| argument-hint | [check | extract <section> | verify] |
Paper Claims
Manages the claims tracking system for a LaTeX manuscript. Three modes based on $ARGUMENTS:
Mode: check (default)
Run when $ARGUMENTS is empty or "check".
Run the full Phase 0 validation suite:
python scripts/check_all.py
If the project doesn't have check_all.py, run individual scripts:
python scripts/check_claims.py --staleness --orphans
Report issues by category:
- STALE: quotes no longer in .tex — update the quote in claims YAML
- SUSPECT: nearby text changed — check if claim still holds
- Warnings: hedge inconsistencies, unresolved refs — fix claims YAML or .tex
- Orphaned provenance: unlinked entries — add provenance_claim refs to claims
Mode: extract
Run when $ARGUMENTS starts with "extract".
- Identify the .tex file and line range for the requested section
- Read provenance.yaml (or equivalent) for evidence references
- Read existing claims files for cross-reference IDs
- Extract claims to
claims/<main|si>/<section>.yaml
Claim schema
_meta:
verified_at_commit: "<current HEAD short hash>"
verified_at_date: "<today>"
tex_source: "<tex filename>"
tex_sections: ["<section names>"]
claims:
- id: <prefix>_<name>
text: "one sentence paraphrase"
type: <empirical|logical|literature|interpretive|scope|qualification>
strength: <strong|moderate|weak|contested>
locations:
- file: <tex file>
quote: "exact 20-80 char substring"
evidence:
- provenance_claim: "substring of provenance entry"
- cite: <bibtex key>
- claim_id: <other claim ID>
- figure: <label>
- table: <label>
supports: [<parent claim IDs>]
depends_on: [<premise claim IDs>]
qualifications: ["caveats in text"]
What is a claim
An assertion a reader could ask "what's the evidence?" — NOT method descriptions, definitions, or transitions.
Strength criteria
- strong: own-data with provenance + replicated independently
- moderate: own-data single dataset, or strong literature
- weak: argument from absence, interpretive, sparse
- contested: evidence for and against
After writing, validate: python scripts/check_claims.py
Mode: verify
Run when $ARGUMENTS is "verify".
python scripts/check_claims.py --staleness — find stale/suspect claims
- Review and fix each flagged claim
python scripts/check_claims.py --mark-verified — advance verified_at_commit to HEAD
python scripts/generate_evidence_graph.py && python scripts/generate_claim_index.py
- Report: total claims, edges, longest chain, weak load-bearing claims
Boundary with provenance
Claims reference provenance entries — they never duplicate source:, script:, or sha256: fields. Use provenance_claim (substring match) and provenance_id (directory name) to link.