Registered before the human's figures: the check stops at 152 items, and its halves are read apart¶

The human checked item-002 to item-153 between 2026-09-24 and 2026-09-26, 152 items past the rubric's worked example. The check stops there, fixed before any human-to-model figure is computed. The page shows items in the batches' seeded order, so the 152 are a random sample of the 383, and a stop chosen after seeing agreement would be optional stopping.

Why 152 and not 383:

• The question the corpus depends on is valid or not, and 152 items estimate it. Per-label specific agreement on the rare labels, and human valid rates per organization (75 and 77 items) and per half (35 to 40 items), are underpowered at this size and are reported as descriptive only. (Corrected after review: this line first said about 19 items a cell.)
• Each lock reveals rater A's label, so the checker is calibrated toward rater A as the check goes on. Later items are less independent of the rater they are compared with, and more of them would raise the agreement for that reason alone.

Two readings added now, before the figures:

• Human-to-rater-A agreement is also reported on the first 76 items and the last 76 apart. A rise from the first half to the second is read as calibration toward rater A, not as agreement.
• Human-to-rater-B agreement is reported on the same 152 items. The human never saw rater B's labels, so this pair carries no reveal.

The checker is the first author, a first-year PhD student who is not a maintainer of either organization's projects. During the check he reported that rater A often read the reviewer's request and the code change more fully than he did. A human-to-model disagreement is therefore not taken to be the model's error by default.