Paper review

Follow research/autonomy/OPERATING_PROTOCOL.md, especially "Review with an endpoint".
Read the outgoing artifacts, existing reviews, and actual publish-bar requirements. Reuse a
completed review when the deliverable digest matches; do not restart because unrelated main
commits moved. An already-reviewed manuscript does not need another baseline round.

Budget one independent review batch, one batched repair, and a focused independent verification
of changed claims and dependencies. Further full review needs a named material reason. If existing
publication evidence requirements remain unsatisfied, report that fact rather than claiming ready.

Apply the operating protocol's scientific follow-through and reassessment before judging a
manuscript ready. State separately whether the analysis is valid and whether its surviving finding
is new and useful enough to merit a paper. Material contrary evidence or substantial claim narrowing
requires a focused contribution decision before further polishing, even if every remaining sentence
is accurate. The minimum repair may be shelving the paper or a bounded decisive scientific check.
Do not rescue a failed contribution with caveats, substitute a positive single model for relevant
coverage, or treat an unavailable measurement as negative. Keep this within the existing review
batch; report the clear verdict, evidence, remaining decision-changing gap and next action.

Require the operating protocol's complete public-EMC evidence evaluation for the actual claim.
Independently check source discovery, eligibility and omissions, including EMC observations within
broader studies, rather than accepting a complete-looking table. No source- or sample-size exception
permits convenience sampling of EMC. Pending suitable accessible evidence blocks manuscript
promotion; verify the reasons and claim consequences for unavailable or unsuitable measurements.

Findings distinguish current submission blockers, maintenance gaps, and optional editorial changes.
Every finding names the exact outgoing claim, source, consequence, and minimum repair. A correct
sentence without a guard is maintenance. Zero findings is a valid result. Verify objections against
the evidence before editing. Keep limitations; replace wrong text without adding defensive padding.

Do not repeatedly mutate the paper while another agent reviews it. Pin the review inputs. Existing
blind-seat and hardening records remain the publication evidence format; do not fabricate or
silently alter past reviews. record_bar_evidence.py records them and publish_bar.py decides
whether all required evidence and authority hold. A review's iteration budget is not permission to
post an unresolved paper.

The legacy reference preserves incident evidence and specialized
procedures. Read only the relevant passage when investigating an actual regression. Its repeated
whole-paper rounds, automatic guard expansion, and retired stopping rules are superseded.