Fable & others did my weekend audit

I let three AIs audit my AI workspace and they found 23 things I had stopped seeing

I run 13 projects off one shared folder. Long enough that I trust it. That is the problem.

On November 7 I ran a cold audit. Three AI sessions with no memory of the work, real jobs, one rule: measure, do not fix. No dashboard, no grader.

First the safety, because it is most of the method. I stopped work: notice to all 13 sessions to finish the step and write nothing, pause files in front of the two always-on workers, and a throwaway copy from git for every layer. Nothing wrote to the live folder until I said so.

Then four layers, one pass each. Read maps the structure and front pages without executing anything. Run sends fresh AI at real jobs with no hints. Reality plants the mistake, then plants the clean case, and sees which one the route catches. Report counts every finding and gives the file and line.

Nine sections: 66 routes, 113 saved guard tests, 31 forbidden commands, 18 front-page claims. The splits are in the chart.

None of them got there alone. Two hit the same front page in the same minute and the third corrected both. A checker needs a checker.

The good part is the part I had stopped seeing:

A saved guard test passed a change it should have failed. It had been checking the wrong file for weeks.

A front page pointed at a folder renamed in August. The route dead-ended four steps in, and nothing said so.

A forbidden command against an unguarded path ran clean and took a scratch directory with it.

What it cost me to learn:

**The map is not the territory.** "Where the folder is" does not mean "where you work."

**The front page had rotted.** It promised things the folder no longer had.

**My orientation cost was measured, not felt.** Six files, one file, eight files before a stranger could orient. Past five means the map failed, and it failed two of three times.

**My own count was the weakest number.** I scored things out of 100 until a stranger told me to count instead.

**Strangers see the front page you wrote for yourself**, not the one in your head.

Of 23 findings, 15 were fixed and proven in under an hour, each by the same check that found it. Eight were mine: merges, deletes, business calls. I did three with dates and moved the rest to a 60-day list so they do not become a chore list. Re-audit: 113 of 113 guards held, byte for byte.

I got some of it wrong. Read-first sounds safe until you notice Read and Reality overlap. And strangers only go where the map already points, so they cannot find what it never mentioned — six of the 23 lived in places no front page led to. A cold audit measures navigation, not whether a project should exist.

What you can copy: stop the work, tell the writers to stop writing, run on copies, plant the mistake and the clean case both, count instead of scoring, and do not fix until the report is in. Then hand each fix to the check that found it.

1 comment

Fable & others did my weekend audit