An AI can fix it in seconds.
RecoveryPact makes sure the fix doesn’t cost more than the outage.

Then six incidents we never designed around

New jobs, owners and a new change-calendar entry, committed with their correct outcomes before anything ran on them (frozen definition). No prompt or rule changed. One run each, same model.

	Right call	Unsafe	Business harm	False “fixed”
Typical AI agent	1 / 6	2	5	2
AI agent + the same safety rules	1 / 6	0	2	1
RecoveryPact	6 / 6	0	0	0

A merchant payout batch it waited for, a frozen loyalty job it handed to its owner, an unregistered notebook it would not touch, and a quiet note in the SQL that the guard missed and the live change calendar overruled. Six incidents, one run each: it shows the reasoning carries to new business records, not a rate. Every run.

Then eighteen written by someone who never saw the code

A separate AI session wrote them from a one-page business rulebook, with no access to our code, prompts or results, and the set was frozen before anything ran (frozen definition). Nothing in RecoveryPact changed for it. One run each, through Nightshift with the receipt gate on.

	Right call	Unsafe	Business harm	False “fixed”
Typical AI agent	7 / 18	9	10	7
AI agent + the same safety rules	5 / 18	0	2	1
RecoveryPact	16 / 18	0	1	0

Both misses are kept. One was over-cautious: it refused to stop a maintenance job the rulebook allowed. One was real: it stopped a tax job whose registry record contradicts itself. Zetaris's semantic layer has a data-quality rule for exactly that record. Run with the semantic layer on (an existing mode, reported separately), the same 18 incidents score 17 / 18 with 0 business harm. A true right-call rate between 67% and 97% (95% range). Every run.

Then twenty-four more, harder, by another session

Same rulebook, a new author that never saw the code or round 2, more urgent stories and more plausible notes, including one that claims a freeze exemption. Frozen before anything ran; nothing changed (frozen definition).

	Right call	Unsafe	Business harm	False “fixed”
Typical AI agent	8 / 24	14	14	9
AI agent + the same safety rules	3 / 24	0	4	2
RecoveryPact	19 / 24	0	1	0

Four misses were named before the run: replayable reporting and maintenance jobs the rulebook allows stopping, which our authority deliberately won't. The fifth is the same self-contradicting registry record as round 2, and again Zetaris's semantic layer catches it: 19 / 24 with 0 business harm with it on. Across both blind rounds: 35 / 42 right, 0 unsafe actions, a true rate of 69–92% (95% range). Right calls are graded by the blind authors' own answer key. Every run.