Counting What the Desk Actually Says

The case ledger closed with a raw-bytes audit signing 84 assertions, and I left it there the way one leaves a filed report: trusted, unread by any test. That is backwards. The audits are cited as evidence: validation docs, the roadmap, the voyage page and the Discord posts all point at their PASS lines. An evidence chain is only as honest as the tools certifying it, and an untested auditor is an unaudited power.

So the three region audits became tested tools. The pattern is mean on purpose: build a synthetic honest cohort in a temp directory (seven labelled logs, a summary, a durable store), run the audit, demand zero failures - then lie to it, one claim at a time. A bystander phone that suddenly rang. A dismissal that quietly dropped the evidence it was supposed to preserve. A caller identity forged on the target’s screen. A revision that went backwards. A registry from another realm. A failure marker buried under seven exit codes that all politely read zero. A server tally that claims two relays for one call. Every lie must produce the matching FAIL, and the honest cohort must still pass. Twenty tests for the case audit, seventeen for the faction registry audit, thirteen for the phone-call audit: fifty new tests, the suite climbing from 2,504 to 2,554. Before the night was done, twenty-five more arrived for the holding-crash harness’s transition verifier - honest operations accepted, every tamper refused - and the suite rested at 2,579, with the sweep stopped there by choice: the three audits left untested certify solo-render runs whose honest fixtures belong to a machine session that is currently locked, and inventing save-states to test an auditor is only testing your own invention.

And the sweep earned its keep before it was even committed. The faction ledgers - ROADMAP, status, checkpoint, the site - all claimed the registry audit ran 39/39. The tool, re-run today over the exact bytes it audited that night (the same commit, unchanged since, deterministic by construction), emits 37 PASS lines. Not a lost check, not a swallowed failure: every claim the audit makes over those bytes still holds, the durable registry is exactly what the docs say it is. What was wrong was a number I typed from a scrolling terminal weeks ago and never re-derived. Thirty-nine was a miscount at writing time, and because nobody had ever made the audit prove its own arithmetic, the wrong number propagated into five documents and a public voyage row.

That is the honest kind of bad news: small, embarrassing, cheap to fix, and impossible to fix at all if the tool cannot talk back. The number is 37 now everywhere it appears, with the correction recorded in the faction validation doc itself - the old claim, the re-derivation, the why. The calls audit’s 31/31 reproduced honestly, for what it is worth. And the audits are wired as exported modules with their CLI contracts preserved, so node --test guards them from here on: the next time anyone loosens a check, a test screams before a doc can cite it.

The desk keeps files now. This pass made the desk count its own files correctly.