A Figure Cannot Police Itself: Auditable Figure Generation for Framework Papers, and 18 Ways the Audit Failed

In a research paper, the data are deposited, the code is released, the statistics are reported with their assumptions, and the prose is contested by referees. The figures are looked at. No mechanism in ordinary practice would catch a number in a figure that has no source anywhere in the project — and framework papers, whose argument lives almost entirely in large, information-dense diagrams, are where that gap is widest. This work presents an auditable figure pipeline: figures generated from machine-readable registries rather than drawn; registries structurally incapable of holding a result, because the columns that would carry one are rejected at load; every figure checked against a registry of retired numerals on its rendered output rather than its source; and every figure checked against every other figure. Applied to a 24-figure framework series across 27 registries, the process retired 44 items with no source (22 fabricated values, 14 cohorts/sites/patients that did not exist, 8 metrics defined nowhere). Most were exposed by the data model rather than by any audit — but two findings do not follow from the count. A fabricated value removed from one figure reappeared, unprompted, in another, as a different quantity under a different caption. And two figures contradicted each other while both passed their own checks: one announced that validation was complete, the other that nothing had been tested. The contribution is not the pipeline. It is the 18 ways the audit itself failed, in 7 classes — satisfied by the artifact it was auditing; checking a constant while the drawing said otherwise; reporting success on finding nothing; rejecting correct work; producing a number that disagreed with itself; and not covering the prose, where a citation was invented. All 18 were found by injection, none by reading the audit code. An audit is an artifact, and it needs its own adversary. Limits. The pipeline proves that no figure reports a number its registries cannot produce, and that no figure contradicts another. It does not prove the registries are right. Consistency is not correctness. This is one pipeline, one author, one figure series: an experience report, not a controlled evaluation. Contents. The manuscript (8 pp.), the paper's own four figures in PNG/PDF/SVG (themselves generated and audited by the pipeline they describe), the failure-mode and audit-surface registries, the audit and cross-figure code, and a single reproducible build gate. Unzip into an empty directory and run python3 verify_all.py. The 24 audited figures of the source series are in the companion TLMM v7.6 record, linked below; the counts in this deposit are frozen with the SHA-256 of every file they were derived from, so they can be re-derived against it.