The self-catches gallery

The exhibits here are my favorite part of the whole project: the times the system caught its own authors — including me, including the AIs that built it. A machine that only catches the market is a backtest. A machine that catches its makers is a governance system. Every entry below is on the internal record with dates and receipts; these are the plain-words versions.

The audit that debugged itself on night one

The first time we ran the self-audit — fifty-one checks, built to fact-check the system against its own claims — it found a real data bug that would have quietly hurt us four days later. Good start. Then it found a second bug: an inverted comparison inside one of its own checks. The auditor’s first catch of the week was the auditor. We fixed both and kept the habit.

The check that corrected its own explanation

We added a governance check that makes sure every numbered decision cited anywhere in the system actually exists in the record — no phantom citations. It passed its first live run, then corrected its own explanatory comment within the hour, because the comment describing the check was itself slightly wrong. The machine proofreads the machine.

The audit that failed the very session that widened it

During the week we expanded the machine to watch basically all of crypto, the nightly governance scan FAILED — on the session’s own paperwork. A queue entry had wrapped across two lines and broke the format the scanner expects. The scanner was right. We fixed the paperwork, recorded the trap, and moved on. It has no idea how big a week it interrupted, and that’s the point: it doesn’t grade effort, it grades the record.

The gate that refused its own school

Everything public on this site passes an automated publication gate — no keys, no configs, no positions, no advice-shaped language. First time we pointed it at the machine’s own teaching library, it refused one of the lessons. The library the gate quotes from failed the gate. So now the publishing side only draws from lessons that pass, and the refused one stays internal. The rule didn’t bend for the house it lives in.

The census that lasted two hours

The day we widened the exchange watchlist, the first census looked clean — until I actually read the receipts window the ruling required and found SHIB excluded by a volume floor at $713k against a $1M line. The easy fix was lowering the floor. The honest fix, which is what we ruled, was admitting the floor itself was the wrong idea and rebuilding the screen as exclusions-only. The census stood for two hours. The receipts outlived it.

That’s the collection so far. Entries get added when the record produces one — no schedule, like everything else here.

— Adam, Studio Amadeus