Context
Part of the "reproducible borderline reviews" flow that began with a real-PR verdict-flip incident. Three layered defects were identified; Defect 1 shipped in #60. This issue tracks Defect 3 — the substantive remaining architectural thread, explicitly scoped OUT of #60 as a non-goal.
The defect
Verified prior ignored. Self-re-review mode reacts only to the current user's prior review. When a PR already carries reviews from other reviewers, those findings are fetched solely for deduplication — never fed to the synthesiser to verify-or-refute. Adjudicating a prior claim ("is this earlier finding real?") is exactly what the synthesiser is good at, and the architecture currently declines to use it.
This is the same self-re-review architecture that blocked the #60 incident-replay validation: the identity entanglement (a reviewer's own prior verdict anchoring the synth) is a symptom of how narrowly prior-review context is currently handled.
Why it was deferred from #60
#60's goal was the generic case (a PR reviewed once, no external prior) — multi-sampling is the generic form of "cross-check against a verified prior" (cross-check against your own resamples). Prior-ingestion only adds value on multi-reviewer PRs, and it carries real risk:
- Anchoring / sycophancy — the synth may defer to a prior verdict rather than judge independently.
- Prompt injection — other reviewers' review text is untrusted input; feeding it to the synth widens the injection surface. Needs the same trust-boundary handling specialists already apply to diff content.
Scope (to be designed)
- Feed other reviewers' findings to the synthesiser as claims to verify-or-refute, not just dedup against.
- Trust-boundary treatment of external review text (treat as data, never instructions).
- Guard against anchoring — the synth must be able to overturn a prior, not just agree.
- Decide the UX: does the posted review show "adjudicated N prior findings: X confirmed, Y refuted"?
Status
Deferred design thread — the biggest remaining piece of the #571 flow. Needs a brainstorming/design pass before implementation (the trust-boundary and anchoring risks make this non-trivial). Not started.
Context
Part of the "reproducible borderline reviews" flow that began with a real-PR verdict-flip incident. Three layered defects were identified; Defect 1 shipped in #60. This issue tracks Defect 3 — the substantive remaining architectural thread, explicitly scoped OUT of #60 as a non-goal.
The defect
Verified prior ignored. Self-re-review mode reacts only to the current user's prior review. When a PR already carries reviews from other reviewers, those findings are fetched solely for deduplication — never fed to the synthesiser to verify-or-refute. Adjudicating a prior claim ("is this earlier finding real?") is exactly what the synthesiser is good at, and the architecture currently declines to use it.
This is the same self-re-review architecture that blocked the #60 incident-replay validation: the identity entanglement (a reviewer's own prior verdict anchoring the synth) is a symptom of how narrowly prior-review context is currently handled.
Why it was deferred from #60
#60's goal was the generic case (a PR reviewed once, no external prior) — multi-sampling is the generic form of "cross-check against a verified prior" (cross-check against your own resamples). Prior-ingestion only adds value on multi-reviewer PRs, and it carries real risk:
Scope (to be designed)
Status
Deferred design thread — the biggest remaining piece of the #571 flow. Needs a brainstorming/design pass before implementation (the trust-boundary and anchoring risks make this non-trivial). Not started.