A Rule Everyone Agreed With, at Corpus Scale
Four lessons back, this track settled something real: a review round that rejects nothing is a failed review, because an empty log certifies an absence of judgment rather than a presence of quality. The rule was measured, it was right, and it shipped unscoped — every round, of every kind, on every pipeline, had to reject something or count as never having happened.
Here is what that costs at scale, and the numbers are worth stating plainly. In the workshop's first two weeks, the gate that checks a review loop closed legitimately was the single most failing check in the system: seventy-nine of one hundred and eighteen verification failures, four times the runner-up. Review chains ran six, eight, twelve rounds. And five quests — including the one you are reading — eventually landed on the recorded override, because no other exit existed.
The Ending the Rule Could Not See
Look at how the honest chains actually ended. The final round would come back as, in substance, a re-check of the repairs: did each fix land, did it break anything nearby. The reviewer granted a pass, the gate opened — and under the unscoped rule that ending was illegitimate, because a round that rejects nothing is a failure. The honest close was being performed as disobedience.
The repair is not a loophole; it is a distinction the rule was missing. Every round is now one of two kinds, and the request says which. A cold sweep is a fresh reader over the full artifact, and there the old rule stands untouched: a cold sweep that reports nothing at all did not read the artifact. A close-out is every round after a repair pass, scoped to the previous round's accepted findings, the repair diff, and the diff's echo sites. Its question is did the repairs land — not what else is out there. A close-out that passes is the loop terminating as designed, and its empty log is the ending, not a skim.
Why do scoped rounds converge where full rounds cannot? Because a full read of a ninety-file artifact is a sample, not a sweep. One chain's second full round found more than its first — twelve findings against seventeen — two readers sampling different neighborhoods of the same city. Fresh full rounds sample new neighborhoods forever. A close-out's scope is the repair diff, and the repair diff shrinks toward zero.
Grades, a Floor, and a Ledger
The rebuild's second half is a way to talk about blockers without treating them all alike. Every blocker carries a grade. B1 — teaches wrong: a factual error against the live source, a mechanism that does not exist, an invented citation, a wrong magnitude. B2 — internally broken: two sites asserting incompatible things, a quiz with a wrong key, an exercise nobody can perform. B3 — blemish: real, recorded, and deliberately below the floor — a blemish lands as repair debt, because repair is reader-triggered and a register slip is not a reason to hold finished work hostage.
The close floor is: the loop closes when no B1 or B2 stands open — each repaired, or declined with a reason on the record, and a decline without a reason is refused where the record is written. The bookkeeping moved out of prose: a finding is a row in a ledger, imported after each round, flipped to repaired or declined one by one. The next round's request is generated from the ledger — repaired findings listed for verification, declined ones carried as do-not-re-raise. What this deletes is the hand-carried tally: late rounds spent on forty-versus-forty-three arithmetic are minutes not spent reading.
Two Instruments, One Ruling
Close-outs are scoped, but repairs echo, and two echo shapes dominate. The first is residue — the retracted claim surviving at a sibling site. Its instrument is a screen that greps the whole quest, both languages, for the retracted claim's distinctive tokens and lists every survivor; each hit is repaired or known. The second is the quiz. Its screen fails a landing only on hard defects — duplicate options, a key outside the options, a key mismatch across languages — and prints the guessing-strategy accuracies as information. The ruling on distributions deserves quoting, because it is the philosophy in one line: out of a hundred quizzes, ninety-nine landing on the same option is still random. A skewed key is not a defect, and these quizzes have no adversarial taker. Distributions inform; they never gate.