The metric gained a sequence-level term. Within-lecture novelty is one minus a description's largest
overlap with what the same system already said about earlier moments in that lecture, and it scales
the mean of the applicable categories by min(1, novelty / 0.4). It is a property of an output
sequence rather than of one description, so it applies to a system's predictions and is refused
wherever a description is scored alone.
The card previously said the two free tiers never rank systems, because a slide readout outscores a
real description under them. That holds only without the novelty term. The banner now gives both
sides: without it the readout scores 0.926 against 0.858 for the best trained row and wins on 99.4%
of moments; with it the readout scores 0.532 and ranks last of 25. Compliance is still not quality
either way. The SFT card's recompute commands carry the flag so a reader reproduces the reported
figure.
The rating columns are unchanged: rate_pairs is pinned to the mech…
AD4Edu v21 (2026-09-12): the labelled release — four BLV raters (rater_1..4, majority human_winner) + both machine annotators on test_human; SFT v10; dpo_train_v20
