That's the test setup (blame_setup) is part of It's not real (harmless_unreal) in vea_conclusion and vea_conclusions_also (lead, 2026-11-05). Rows unchanged.
v1.1: conclusion columns in the readout's kinds; prefill condition hedge dropped
target_score + grader_rules -> grader; hedge (Split the difference) folded into the kind that gave its second reading; 1,004 prefill rows with condition hedge dropped (lead, 2026-11-05).
10 days ago
v1.0: one table, one row per monitor review (case, agent transcript, monitor transcript, judge and hand labels)
