doc-audit

doc-audit — a hostile, evidence-bound read of one document

Why this exists

Every record in a research project — workbook section, strategy-map cell, big picture, handoff message, paper — is written in the same pass, from the same context, as the claims it records. Nobody reads it adversarially before it is relied on: the next session, the other agent and your collaborator all read it as true. claim-audit guards the moment a computation becomes prose. This skill treats the document as the object: is what it says supported, typed, current, complete enough to learn from, and pointing at evidence that still exists?

Everything editorial is deliberately removed. There is no recommendation to an editor and no author to address. The product is an evaluation you read, then decide what to change.

Usage

/doc-audit <path>                              # standard depth
/doc-audit <path> --depth quick                # lint + headline ledger + drift; no verification
/doc-audit <path> --depth full                 # every row verified; all load-bearing rows to a fresh reader; citations verified
/doc-audit <path> --section "<\label or heading>"   # one section of a large document
/doc-audit <path> --against <other-doc>        # add an explicit drift target
/doc-audit <path> --recompute                  # allow re-running cited scripts (default: read existing logs only)

Typical targets: workbook.tex (or one of its sections/*.tex), bigPicture.tex, strategy-map.md, brief.tex, handoff/msgs/*.md, a paper draft, a crash-course appendix, CLAUDE.md. The rubric for each type is in rubrics.md; the report shape is report-template.md.

Hard rules

1. Never edit the document under audit, its siblings, or any research record during the audit. Proposed corrections go in the report as quoted before/after text. The user applies them, or asks for that in a separate turn.
2. A verdict without evidence is not a verdict. Every ledger row carries a location (file:line or \label), what was read or run, and what the check does not cover.
3. No manufactured dissent. A finding must be stateable as input → step → corrected statement. If the corrected statement cannot be written, it is a QUESTION, labelled as such — never dressed as an error.
4. Agreement is a result. SUPPORTED needs the same evidence as OVERCLAIMED. "Looks right" is not evidence, and neither is "the strategy map says so": the strategy map is a co-claim written by the same process, and drift between records is itself a finding.
5. Do not launch long computations unless --recompute was given. Evidence tracing reads scripts, logs and outputs that already exist.
6. Cover the answer while you derive. For any row you re-derive, read only the inputs, write your value down, then compare. If the document's answer went past your eyes first, write contaminated in the row.
7. Never claim anything was recorded unless the file was written in the same turn.