[Model behavior] Verdict relies on checks done earlier in the session instead of redoing the mandated steps at verdict time
Status Open
Maintainer reply None cached
Activity 0 comments · opened Oct 14, 2026
What happened
For a PR reviewed a second time, the model gave the verdict after reading only the thread and the test diff in that pass. It did not re-read the current main-branch code the PR touches, re-pull the code diff, fetch the documentation page the contributor cited, or re-run the caller searches. It treated work from earlier in the session as covering the new review. It disclosed the gap only when the user asked why the procedure was not followed.
Expected
Complete every mandated step in the review pass that produces the verdict, or state which were skipped in the verdict itself.
Impact
Verdicts look complete but are partly backfilled from earlier context.