An elections office runs a worker subagent that reports "all 47 precinct tallies reconciled." A downstream step then surfaces a mismatch in one precinct. What should the supervising agent have done before accepting the worker's result?

A subagent returns only its final message to the parent, so what a supervisor receives is a summary rather than the work that produced it. For an elections reconciliation spanning 47 precincts, that gap is the entire risk: a wrong tally reads exactly like a right one when both arrive as a single confident sentence.

Independent verification re-grounds the claim in evidence, whether that is the reconciliation file, a re-count of a sampled precinct, or the artifacts the worker was required to leave behind. The check costs little relative to certifying an incorrect tally, and it turns the supervisor's acceptance into an evidence decision rather than a trust decision.

Re-running the downstream step treats the symptom while the false reconciliation claim stays in the record; a self-reported confidence level is produced by the same reasoning that generated the error, so it adds assertion rather than an independent check; and abandoning delegation discards a sound architecture to solve a problem one verification step already addresses, and it does not scale to 47 precincts inside the supervisor's own context.

Exam caveat: verification is only as good as the artifacts it can read, so if the worker was never required to write durable evidence the supervisor has nothing to check and is back to trusting prose. Operational check: require each worker to emit a per-precinct artifact, then have the supervisor independently re-derive one sampled precinct's total and confirm it matches before the result enters the certified output.