Currently thresholds are heuristic: N<20→0.40, N<50→0.25, N<100→0.15, ≥100→0.10. With 11 Atlas entries we can start back-fitting these from real verdict distributions.
Acceptance: replay each Atlas entry's baseline distribution, derive 90th-percentile thresholds from random-direction-real-acts null distribution, ship as v0.2.0 default with old thresholds reachable via legacy_thresholds=True.
Currently thresholds are heuristic: N<20→0.40, N<50→0.25, N<100→0.15, ≥100→0.10. With 11 Atlas entries we can start back-fitting these from real verdict distributions.
Acceptance: replay each Atlas entry's baseline distribution, derive 90th-percentile thresholds from random-direction-real-acts null distribution, ship as v0.2.0 default with old thresholds reachable via
legacy_thresholds=True.