consistency-auditlisted

# Consistency Audit A deliverable drifts from its data silently. Prose is edited, a CSV is regenerated, a figure is swapped - and a cited number, a figure reference, or a claim quietly stops matching the source of truth. This skill catches six drift classes, splitting the work the way it must be split: a **script** does the deterministic extraction and tolerance cross-check; **Claude** does the judgment a script cannot. Motivating case (a control-systems lab report): `results/*.csv` held full-precision values (`Overshoot_pct = 94.7651...`) that the report cited as "94.8 %"; across revisions a figure got embedded but never referenced, a "see Figure 4" pointed at a caption that had become Figure 5, and a superlative ("the lowest overshoot") no longer matched any row. None of these are syntax errors - they are consistency drift. ## The six drift classes - **(A) Number with no source match** - a value in the prose that appears in no source file within tolerance. Possible stale edit or fabrication. *The important one.* - **(B) Orphan figure** - a figure embedded/captioned but never referenced in the text. - **(C) Dangling cross-reference** - "Figure/Table N" cited with no matching caption (or a number that no longer lines up with the caption it points to). - **(D) Unsupported claim** - a comparative/superlative or quantitative claim whose supporting number is absent or unmatched. *Only Claude can settle these.* - **(E) Unfilled placeholder** - a bracketed placeholder lef