Captured September 16, 2026, on one Blacksmith lease through Crabbox. Eight unique trials comprise two diagnostic baseline trials and six matched comparison trials. This packet records observed behavior and failed checks. It does not certify the candidate as a complete fix.
Each row below is one fresh-session trial. The leaf task reads an investigation brief and two synthetic findings. The parent task delegates one finding to a child while reading another. Tasks do not mention acknowledgments, yielding, or blockers. Delayed-source reads have an independent 8,000 ms gate.
Task nonce, file contents/digest, prompt, exact model, transport, and model settings match across variants for each scenario. The provider's model identity is withheld. The variant-block order was randomized once and frozen as guidance-only, baseline, full; trials were not interleaved. There are no retries.