Why “done” has to mean verified

The failure mode that kills trust in automation is not the wrong action — it is the confident wrong report. A refund that half-executed. A digest that missed a theme. An agent that says done when it means probably.

Every Turbine run ends in the same three steps, no exceptions: reconcile, replay, then report. Reconcile means the ledger delta, the ticket count, the API response — some external number must agree with the claimed result.

An unverified “done” is not a result — it is a hope with a timestamp.

Replay means the checklist runs again against the finished work: did every planned step complete, in order, with the expected artifacts? Drift between plan and output stops the run.

Only then does the run report done — with the receipts pinned. If either check fails, the run flips to needs-review and a human gets the evidence, not a shrug.

This costs us seconds per run. It is the entire reason operators stop checking Turbine’s work after week four — the checking is built in.