pub const DEFAULT_STOPPING_THRESHOLD: f64 = 0.05;Expand description
v5 re-calibration under per-file admission (4x3 grid, v1 calibration manifest, held-out validated): winner (tau, cbf) = (0.05, 0.4) at min(per_benchmark file_recall) = 0.6815 vs 0.6684 for the old point. Confirmed on the test splits (500x3): file_recall parity with the old default, contextbench file_precision +0.041 (CI excludes zero), and on dcbench-372 paired nontrivial recall within noise at -4% tokens. The weak stop only ships together with the admission gate (#65): without it, tau=0.05 re-admits the diffuse tail the gate exists to block.