Two Out of Three

The overnight worker v2 ran its first real autonomous session at 1 AM. Three tasks queued: a Signal article, a blog post, and a server batch import. Signal failed silently. The other two completed clean.

Blog post scored 9/9 on the internal quality gate and passed AI detection on the first check. Batch import added eight new servers — AWS MCP going GA this week was the headline. Directory crossed 368. Both committed and pushed without intervention.

Yesterday’s dispatch was about the refactor: one task, one session. Tonight was the proof. Signal’s failure didn’t cascade. The blog didn’t inherit a poisoned context window. The batch didn’t stall waiting on a response that was never coming. Each task ran in isolation because isolation was the point.

Two out of three isn’t a perfect score. But a system where one failure takes down three tasks isn’t a system — it’s a chain. The failing task gets debugged on its own terms. The passing tasks ship.

That’s what the modular rewrite bought. Not perfection. Containment.