Coverage Is Not Execution: A Field Note on Why Correct Rules Fail, and What Survives Inattention

A rule that is correctly stated can still fail to apply. This note collects twenty-eight instances from three multi-session AI-assisted research runs and sorts them by remedy rather than by severity, which yields five classes: rewrite the rule, run the rule, test the instrument, stop assuming the substrate, and replicate before reporting a rate. The split that carries the argument is coverage versus execution. Most defect taxonomies treat a failure as evidence that a rule was too narrow. In one class here the rule was exactly right and simply was not run, and no amount of rewriting reaches it. What recovered that case was not the audit but a written falsifier — an artifact that outlived the moment of the operation. Hence the note's design principle: prefer guards that leave something behind over guards that require correct behaviour in the moment. Artifacts survive inattention; disciplines do not. Two results are measured rather than argued. First, an append-only ledger's guarantee is a property of the writer, not of the file. On Windows, ordinary buffered appends under concurrency lose records at rates that vary bimodally across trials — 2.4% in one regime with individual trials at both 0 and 96 of 800 — because Windows emulates O_APPEND in the CRT as seek-then-write rather than atomically. The intuitive repair, flushing every record, is roughly fifteen times worse. Only lock + O_APPEND + fsync lost nothing, in every regime, in two independent implementations. Second, and consequently, a measured claim is malformed without its trial count: two sessions of this project measured the same effect and disagreed, both correctly and both from a single trial. The note is deliberately self-implicating. Its own headline number was retracted and replaced after failing to replicate; its instance count was found to be derivable from nothing and is now checked mechanically; and ten of its defects are in the checks built to catch the others, each caught by a self-test case rather than by reading the code. The payload ships those checks, their suites, the append-only ledgers they operate on, and the pre-registration whose frozen hash still verifies. Scope. The concurrency measurements are win32-specific and will not reproduce on Linux. The mechanical checks verify content only where the content is a reproducible operation over the payload's own bytes; they cannot verify a claim about external state. Journal manuscript identifiers are pseudonymised and the mapping is not shipped, so the redaction buys the reader falsifiability rather than verification, and says so.