The speed work hit your numbers, with receipts. What is still open is whether the gate can be trusted about its own coverage — one adversary now says the work holds, and the other’s three findings stand unanswered.
This is not the open question. It was measured tonight on real runs, and anyone can recompute it from the run records.
Two independent adversaries attacked the same thing: the part that decides what gets verified. They did not agree, and the disagreement is the most useful thing on this page. Nothing merged. Trunk was never touched.
Any check's command can be replaced with “do nothing,” and everything still passes — because the check that would object is the very check that got switched off. Its objection only runs if it is already running.
Links pretending to be pages were cured. A pointer to a whole external repository, sitting at a content-shaped path, was not. Same class, uncured.
Recording “nothing guards this, and here's why” requires only that someone wrote words. No evidence. Delete the check, its entry, and the excuse together, re-generate, and the gate reports fine.
And the one neither adversary could close: each check declares what it protects, and nothing checks whether that declaration is true. One check claims to guard 11,721 files. The completeness check verifies a declaration exists — never that the check catches anything inside what it claims. The red team ran five attacks at it, got five passes, and refused to call them breaks because its own control never went red. That surface is open, not cleared.
One more, named as a trade rather than a break: the check that catches machinery drift used to run on both lanes — the backstop for the classifier mislabeling something. It now runs on one. The classifier is tighter than it was, so this may be right. Nobody wrote it down as a decision.
Second lens (codex-verdict), all three probes RED. (1) The registry accepts command: "true" for all five load-bearing checks and the rendered workflow still matches its hash. (2) A gitlink (mode 160000) at a genome-matching path classifies as genome-data; mode 120000 symlinks were cured, gitlinks were not. (3) Recorded-absence accepts arbitrary prose with no check-guard evidence; synchronized deletion of file + entry + absence then re-mint yields a green gate.
Falsifier (opus): final verdict CLAIM-HOLDS for the delta — 23 of 23 machinery checks and 4 of 4 content checks green, run the way CI runs them. Every break it confirmed in round 1 is fixed. It corrected itself twice along the way, and both corrections matter more than its verdict: (1) its round-1 headline — the “production redirect hijack” — was its own harness corrupting a trailing newline via .trim(); re-run byte-verified, that check exits 0 on the plant at both heads, so it never guarded redirect-target rewrites at all. (2) Its two reds this round were its own clone sitting on a detached commit while the real workflow creates a named branch first. Chasing that down exposed a genuine round-1 break it had certified as GREEN: the rebuild had dropped the branch-creation step, and two checks would have failed in CI. It says so plainly: its instrument diverged from CI in exactly the way it was hired to catch in others. Coverage explicitly incomplete on the declaration surface.
Keeper-verified first-person: the content lane is now four checks; nineteen declare they guard nothing. The byte-identity binding is real (mutating the pinned Node version yields a drift error with both fingerprints and a named cure) — but it proves the bytes match, not that the commands do work. Also recorded: the keeper's own first probe of that binding was vacuous, mutating a string that no longer exists in the file. Caught by checking the sabotage actually landed.
Every cure so far verifies form, never effect:
Not one of them demonstrates that the check can catch anything. Each round moved the claim one level up and left the new level unverified. Patching the three live findings would produce a fourth level with exactly the same property.
Your constitution already rules this case: a second refutation of the same surface indicts the frame, not the fix — never accept a third cleverer patch. Two rounds of this surface have now been refuted, and the second round's findings are unanswered.
Tonight also produced the argument in miniature, from a source that costs nothing to trust: three separate instruments reported greens that were not real. The red team certified a genuine break as green because its clone differed from the real gate. Its headline finding was its own whitespace bug. And the keeper's first attempt to test the new binding passed while breaking nothing at all. None of those were carelessness — they are what happens when a green is believed without being shown able to go red. That is precisely what the rule below makes impossible to skip.
Every check must carry, and pass, a proof that plants a violation of what that check claims to protect, runs it through the exact command the real gate runs, watches it go red, then restores it and watches it go green.
What that single rule closes, without one patch per finding:
This is your own law — a check that cannot fail proves nothing — applied to every entry mechanically, instead of depending on anyone's diligence. It is the last rung: there is nothing above “every claim is demonstrated by a proof that itself runs.”
Every check earns its place with a demonstrated failure. There are about twenty-two. Most proofs are small; a few are genuinely fiddly. The proofs cost lap time, but they are cheap next to the checks themselves, and they run in the heavy battery — which is paid once per machinery change, never per teammate edit.
It is real scope growth on this arc that was not in the roadmap. That is the whole reason this is in front of you rather than already building.
This arc grows by roughly a day. The port then carries a proof-backed gate into the factory, and every practice born afterwards inherits a gate that cannot lie about its own coverage.
Merge the speed work now, take the fifteen-minute battery and the sub-minute edit lane, and open the trust question as its own arc later.
Today's gate lists its checks directly in the file that runs them, and its anti-rot guards catch edits to that file. The rebuilt gate moves those commands into a list that currently accepts do-nothings. Merging B as it stands trades a trust property away for the speed. A middle path exists — patch the three named findings, merge, defer the rule — but that is precisely the third cleverer patch, and it is named here only so the option is visible, not because it is recommended.
The red team's final verdict is that the rebuilt gate holds — every break it confirmed is fixed, and it found no new one. The full battery is green and lands in fifteen minutes. So B is a real option now, not a consolation prize. What keeps the keeper on A is that the three findings from the other adversary were never tested by the red team, and the first of them — a check that can be switched off without anything noticing — is the kind of hole this whole arc exists to close.
Build the proof-backed gate, or bank the speed and defer? Everything else on the road is unblocked either way — today's gate still works and still tells the truth; it is only slow.
The keeper reads it as a real loss of a trust property. If you read it as small, that changes the recommendation honestly.
It is written to be the last rung. If it is over-built for what you want, this is the moment to say so — before it becomes twenty-two proofs.