AI Systems Auditing & Governance: The Jerry Ledger

A governance and auditing layer for autonomous AI systems, built across a 12-lane AI development factory, an autonomous trading bot, and a software-building engine. The through-line: make systems demonstrate their claims instead of asserting them. Verification against reality, not documentation.

Every mistake, near-miss, or correction in any lane gets a name, a timestamp, and a permanent record in an append-only ledger. Same session, no exceptions. A mistake that only gets "remembered" is itself a violation: a logged-but-not-ruled Jerry is a Jerry.

90 entries since mid-June, 42 already turned into enforced rules with real detectors. Notable catches:

A lane claimed a code edit had landed on the main branch. It hadn't, a scripting quirk wrote it to the wrong working copy. Standing rule: verify the diff landed before claiming it did.

A governance gate meant to block bad commits read a file path that doesn't exist in the environment it protects. It would have silently approved everything forever, indistinguishable from a gate that was never needed.

One lane filed an urgent complaint against another for missing a 3-day deadline while sitting on two unfulfilled promises at 20 and 37 days. The ledger caught its own hypocrisy.

The same task was dispatched three times in one session because nothing checked whether it had already shipped.

None of these are hidden. All 90 are on disk, timestamped, and half are load-bearing rules instead of stories nobody checks twice.

Systems that demonstrate their claims instead of asserting them. Verification against reality, not documentation. Failure-class taxonomy. Enforcement design.

Claude Desktop was becoming unresponsive on my machine, repeatedly, mid-session. I don't trust that kind of explanation, so I measured instead.

An AI coding tool kept forcing its most expensive model onto routine, mechanical work, no matter what I did to stop it. This is the diagnostic.