Enrollment is not coverage: integrity gates, their measured detection rates, and the limits of self-checking documentation

Abstract

This is an experience report with n = 1, self-reported: the same author wrote the documents, the gates that check them, and this account of what the gates found, and no independent party re-ran anything. Every structural failure it describes was found by the apparatus catching itself, which is the weakest form of evidence that an apparatus works. What distinguishes it from an essay is only that the fixtures and the measured detection rates are published with it, in a form a reader can execute. Abstract A document-plus-code package was audited across some two dozen rounds by an external reader. Each round found real defects. In response the package grew an apparatus of automated gates, written into manifest.py, which check the documents against the code and against each other: that every stated number reproduces from a named generator or matches the tree, that every cited path resolves, that every artefact has a producing script, and — later, and more interestingly — that the claims the documents make about the gates are true. The headline result is negative and worth stating plainly: every defect the apparatus found was in the description of the work, and the physics underneath held every time it was checked. The mathematics did not move between the first round and the last. What moved, repeatedly, was the connective tissue between representations of the same quantity. The numbered sections that follow are about that connective tissue. They are not specific to zeta, to physics, or to this package; they apply to any corpus large enough that its description is itself an artefact requiring maintenance. If you adopt one thing from this note, adopt the cheapest: when your own prose trips a check, rewrite the prose rather than widen the exception. Every exclusion is permanent and grows silently; every rewrite is local and costs nothing but a sentence. The reflex runs the other way, because an exclusion is one line and a rewrite is thought, and the exclusion always looks like the smaller change at the moment you make it. It is not: it is the only one of the two that can hide a future defect. The rule and its single exception — a form that is the evidence cannot be reworded without destroying it — are stated in the note. What is deposited CHECKING.md is the note. VERIFICATION.md is its evidence appendix: the gate-by-gate record, the errata ledger that supplies the replay fixtures, and the measured detection rates printed by the run rather than asserted. The single worked case is the companion record, a package realising the Riemann zeta zeros as a log-frequency comb in acoustic, optical and frequency-domain form. Every measurement quoted here comes from that package's gates and its errata ledger. Relation to the companion record This is Unit B of a two-part deposit. Unit A is the worked case and carries the code; this unit is the argument about it. The relation is recorded as isSupplementTo Unit A's version identifier. At the moment this text was written that relation did not exist, deliberately. Unit A's identifier is not minted until Unit A is published, and a record that ships a relation to an identifier which does not yet resolve is a dangling citation in its own metadata — which is an entry in this package's own errata ledger, and would have been that same defect committed in the file a stranger reads first.