The Same Inversion, One Level Up
Start from the base case, stated in one line: given how reliably a first draft contains something worth arguing with, zero findings has two explanations, and the far likelier one is that nobody looked — so an empty log certifies an absence of judgment rather than a presence of quality. That much is settled, and the case for it is made at length in Beacon Quest, for generated assets. What this lesson adds is what happens when the unit being judged is not an asset but a whole round — one reviewer, one artifact, one verdict covering everything at once.
The inversion survives the change of scale and gets sharper. A per-asset judgment that comes back empty is one skipped look. A review round that comes back empty is a claim about an entire body of work, made in a single sentence, by somebody who had no obligation to show you where they looked.
So the requirement changes shape too. For an asset it is enough to demand a verdict with a reason. For a round you have to demand something a reason cannot fake: a location. A finding that names a file and a field proves the reader was there. A finding that says the argument is sound proves nothing at all, and reads exactly like a round that never ran.
What the Claim Is, Precisely
State it narrowly, because the strong version is indefensible. The protocol does not claim the reviewer's judgment is correct, and it does not claim their taste is good. The claim is exactly: nothing un-judged ships.
At round scale that has a specific consequence worth naming. The reviewer is allowed to be wrong about every single finding and the round still did its job, because the author then has to argue back with evidence — and arguing back is a second reading of the same material by somebody who now has a reason to look closely. A round that produces four wrong findings has still caused the artifact to be read twice. A round that produces none has caused nothing.
That the claim is weak is what makes it defensible rather than what makes it cheap. "Our reviewers have good taste" cannot be demonstrated, so the argument degenerates into a contest of taste. "Every finding in this round names a location a second reader can go and check" can be counted, and when it is not true you can point at exactly where.
Rounds Do Not Simply Go Down
A healthy sequence does not fall cleanly, and a workshop that expects it to will stop too early. Counts oscillate, because a round samples the artifact rather than sweeping it and because a repair seeds defects of its own. What a round tells you is what it named and whether a second reader could act on it — never how many things it named.
Two shapes are worth recognizing. A round whose count rises usually means the reviewer opened a drawer nobody had looked in before — a different slot, a different field, a different layer — and the spike is new coverage rather than regression. And the round that fixes the last finding is not the last round, because repairs introduce their own defects: one measured sequence, run over more rounds than the trend below, had three of its ten findings introduced by the previous round's fix, all sitting in the same file as the fix, in prose the diff shows as unchanged.