Selective rigor in agentic software reconstruction

thesis-review · examiner-tested

An anonymized case study of a frontier model commissioned to audit and rebuild a single-developer system, then run past two independent adversarial examiners who both returned major revisions, conceded point by point.

Findings

• The model closed every P0–P1 defect but declined the structural refactors its own audit prescribed.
• A single-case hypothesis named a “structural patience deficit” on long, gate-less work.
• Two independent adversarial examiners both returned major revisions — conceded point by point rather than defended.