## THE HOOK

KPMG pulled their AI-usage report this week after apparent hallucinations made the numbers unreliable. A big firm, real researchers, still got burned by output they didn’t verify. I read that and felt it. I’m arming an overnight orchestrator on my own business system right now. The difference between “this helps” and “this quietly makes things up” is entirely in how honest you force the system to be when it reports back to you.

## WHAT SHIPPED

Yesterday I armed and installed the overnight orchestrator. It’s wired to run AUTO-OK work only — a narrowly scoped class of operations that don’t touch anything requiring my judgment. It runs on a branch, not main. Default is off. That’s not laziness; that’s the architecture. A flag you have to explicitly set to True is a gate you have to consciously walk through every single time.

The other half of the commit was making the morning report honest and reviewable. This matters more than the orchestrator itself. If the overnight run produces a report I can’t audit — vague summaries, no counts, no diff of what changed — then I have no idea what actually happened while I was asleep. The report has to show me exactly what ran, what it touched, and what it skipped. Anything it couldn’t handle confidently should surface as a flag, not disappear into a success message.

The skipped items and the flags are the whole point. An honest “I didn’t do this” is worth ten confident wrong answers. That’s the KPMG lesson, applied at 2 a.m. on a single-operator business.

## THE LOG