The audit after the last step: three independent cells read the manuals, the gates and the live edge — what they confirmed, what they broke, and what waits on a ruling

The operator's order was explicit: the audit is the last step, it runs only after everything else is done, and its cells may read every file in the project folder if they need to see what is at the bytes. Three cells ran in parallel on disjoint questions, each flag-only, each under the same guards the earlier campaigns carried — no organism woken, no secret opened, no sibling name written, no figure lowered without the register. Cell A re-derived every grade blind against the criteria file's band rows and the eight manuals. Cell B attacked the test suite, the hermetic deploy and the leak gate's newest rule for false greens, weakened guards and self-serving verdicts. Cell C fetched the live worker's surfaces and the shipped bundle on disk and swept them with three disjoint instruments.

Cell A confirmed the Security hold and its printed reason, confirmed Design and Guardian-ARM on the live measurement, and disputed three of the night's moves: it would set AI Systems Engineering at B, Innovation at B+, and Software Engineering at B+, and it caught one false premise in the whole's headline — the claim that Strong was cleared everywhere while one dimension's own text said its Strong row was unmet. The three grades went to the register with both values, the row, the evidence and the investigation, and the operator ruled the same night: all three stand, no downgrade, register closed — the objections name a statement still to be written and introductions days from their evidence, not a claim that fails. The builder's answer is on the register: agree on Innovation, agree on AI Systems on a different ground (the criteria file's own axis for LLM-routed systems, now declared under the bar), and agree to hold Software Engineering with the caveat that it is the grade resting most on the introductions of the coming week. The headline was corrected the same hour, and AI Systems now states the axis its grade rests on rather than a row it does not clear.

Cell B found what an author cannot see in their own guard. The leak gate's build-root self-reference rule, written the night before, used a one-character class where a Windows path needs a backslash, so on this machine the root read as neutral and a self-reference in the shipped tree was never fatal on a main-tree run — a broken guard shadowed by a second control, the hermetic root, which is why nothing leaked. The rule now folds every path to one form before it judges it, judges the enclosing path token rather than a fragment, and the bite-proof plants against it: a self-reference under .next stays green, the same string under .open-next with an operator root goes red. The Worker entry and its Durable-Object modules, which wrangler uploads and the gate never opened, are inside the scan; the path classes now read the server bundle too, with three vendored documentation placeholders named whole rather than waved through as a class; the exact reader-table layer runs at deploy time; the assessment suite pins every published figure so a silent change cannot ship green; a record source with a line number that matches no anchor shape is a failure rather than a skip; a perf floor cannot exceed its measurement plus fifteen per cent. Cell C found no sealed name, no operator path and no secret on the live edge or in the shipped bundle, and confirmed live equals shipped by build id and chunk digest. The cures were re-proven by the gate's own self-test, twenty-six scenarios red and green, and by running the corrected gate against the hermetic build that is live.