Builds merged and their gates

Whole-turn diff per corpus

Nodes and documents are over the documents both runs parsed without error (run_diff's scope). Run totals below count every document in each run directory.

Run totals

10-Q tier presence

Text-corpus crosswalk

Random audit, re-scored

Transitions: what the new runs did to the judged windows

Recall gap by rejection reason

External agreement, before and after

Build a review set

Select windows from the window index (runs/viewer/windows.parquet, every judge bank) by its columns and by the documents a Turn 12 gate touched, take them all or a seeded sample, and save the list as an immutable set under gold/review/sets/. The same seed and query always give the same list. A saved set opens in the viewer at /?set=<slug>.

Documents a gate touched

A window matches when its document (accession, and sequence for exhibits) is listed in a chosen artifact under runs/judge/. Several artifacts are OR; a loss class or kind narrows to documents with at least one loss of that class or kind.

Save

Saved sets