Revise the Artifact, Not the Score

Review creates pressure to optimize for the reviewer. That is useful only when the review points at the task contract. If the author starts changing prose or code merely to make a model stop commenting, the artifact becomes tuned to one instrument rather than improved for its destination.

A valid repair changes the artifact and reruns the evidence that exposed the defect. It then receives fresh review because the new change may introduce a different issue. Editing after the last review while keeping the old verdict is provenance drift: the reviewed object no longer exists.

Do not count findings as a quality score. Ten precise blockers in a large artifact may produce a stronger result than one vague comment in a small one. Counts help reconcile dispositions; they do not rank authors, brains, or artifacts.

Repairs should stay scoped. A reviewer may notice adjacent debt. If it is not required to satisfy the approved gate, record it as an observation or a new task. Turning every review into cleanup broadens authority and makes the landing impossible to reason about.

Repair the Reproduction

For each accepted blocker, keep the smallest reproducer, make the scoped change, rerun the reproducer, and note the result. Then ask the independent reviewer to inspect the current artifact, not a diff summary written by the author.

A reconciliation table can count raised, repaired, stetted, and remaining findings. Its purpose is completeness: every claim has a disposition. It must never become a leaderboard.

A repair is complete only when the artifact tells the truth. Change the smallest responsible surface, rerun the evidence that exposed the defect, and ask a fresh reviewer to inspect the result. Negotiating the count downward without changing the work merely edits the scoreboard.

Code

Reconcile finding dispositions without scoring quality·python

findings = [
    {"id": "R1", "status": "repaired"},
    {"id": "R2", "status": "stetted"},
    {"id": "R3", "status": "repaired"},
]
allowed = {"repaired", "stetted"}
remaining = [f for f in findings if f["status"] not in allowed]
assert not remaining
print({s: sum(f["status"] == s for f in findings) for s in allowed})

Revise the Artifact, Not the Score