Why: The approved v7 sample had two committed checkpoints but failed semantic and final-result acceptance. Evidence: Import the original failing SELF receipt, bind independently replayed usage/source findings, and pin the corrected read-only evaluation. Impact: Keep model quality and merge gates closed while preserving six draft PRs for review. Refs: Agent-System-Dev#512; Agent-System-Backend#193; Agent-System-Documents#95
ui
@ 3b34dfd0c6