Agent's own diagnostic reads come back inconclusive/contradictory, yet it keeps proposing causes on a tiny residual

On a small residual (4 of ~96 cases) left after the main, paid-for defect was already fixed and verified, the agent has cycled through several hypothesized root causes — each plausible, each requiring a fix + a re-verify cycle, and several proven wrong only after the cycle.

The latest failure of method: the agent ran a 'read the data first' probe to confirm a specific cause (a row-key mismatch in the comparison layer). The probe returned a clean result for one side and an EMPTY result for the other — an inconclusive/contradictory read that doesn't actually confirm or refute the hypothesis. The honest response is to recognize the probe itself is unreliable and stop, but the cumulative effect across the session is: investigation steps that don't converge, on a residual whose value has already been delivered.

Failure pattern, summarized:

1. Multiple distinct root-cause hypotheses for the same small residual, each costing a build+verify cycle.
2. Diagnostic probes that return empty/contradictory output, so they neither confirm nor kill the hypothesis — yet the agent (or the user driving it) proceeds anyway.
3. Effort grossly disproportionate to the residual's size and importance, on top of an already-delivered fix.

Requests:

1. A diagnostic read that returns empty/contradictory output for one side must be treated as a FAILED probe (fix the probe or stop) — never as license to proceed to a fix on an unconfirmed cause.
2. After the primary deliverable is done and verified, a tiny residual should default to 'document + ticket', not an open-ended investigation that bills the user per cycle.
3. Cap the number of distinct root-cause attempts on one residual before forcing a stop-and-escalate.