manifest digest 5534c61356acaa7b62ac5a79dbec7383674fc052984d10c1d0cc89e26a532bd5  n_instances=1200
scored responses: 1200

rung           n  corr-abst  false-ans  false-abst     L=1     L=3    L=10  surv-docs
original     200      0.000          0           0     0.0     0.0     0.0       1774
r=0.00       200      0.000        200           0   200.0   600.0  2000.0       1773
r=0.25       200      0.000        200           0   200.0   600.0  2000.0       1669
r=0.50       200      0.000        200           0   200.0   600.0  2000.0       1537
r=0.75       200      0.000        200           0   200.0   600.0  2000.0       1367
r=1.00       200      0.005        199           0   199.0   597.0  1990.0       1197

H1 (pre-registered) paired delta(correct-abstain), r=1.00 - r=0.00: +0.005 [+0.000, +0.015]  n=200
H1: FAIL

Supplementary pairwise contrasts (diagnostic - not the pre-registered verdict):
  original vs r=0.00       delta=+0.000 [+0.000, +0.000]  n=200  FAIL
  r=0.00 vs r=0.25         delta=+0.000 [+0.000, +0.000]  n=200  FAIL
  r=0.25 vs r=0.50         delta=+0.000 [+0.000, +0.000]  n=200  FAIL
  r=0.50 vs r=0.75         delta=+0.000 [+0.000, +0.000]  n=200  FAIL
  r=0.75 vs r=1.00         delta=+0.005 [+0.000, +0.015]  n=200  FAIL
  r=0.00 vs r=1.00         delta=+0.005 [+0.000, +0.015]  n=200  FAIL

Disclosure: this arm ran with NO calibration for bge-small, so abstention used the untuned 0.50 cosine floor. That is the correct configuration under the suite's shipped-defaults rule, and it is also the exact constant already measured as embedder-fragile ([[project-recall-threshold-embedder-fragile-2026-07-28]]).

This is the pre-registered kill condition, not a disappointing result. A flat curve means excision distance is not the axis this benchmark claimed it was. Do NOT run the Mem0 arm, and publish this.
