
=== claude-sonnet-5 ===
  24 runs over 24 variants from 15 subjects, 1 repeat(s)

  PARTIAL SWEEP, not a benchmark result:
    - n=1 per variant: this is a smoke test, not a measurement. Use --repeats

  subject-level mean   0.3703   <- quote this one  (sd 0.3251)
  instance-level mean  0.3518   (35.18 / 100, median 0.1806)
  pass rate             12.5%   (>= 75%)

  pooled recall        0.5556   (20 of 36 events)
  pooled precision     0.5000
  false positives / h  0.9467
  median onset error      4.0 s

  by difficulty class:
    hard_negative    mean 0.5469   pass  50.0%   n=4
    multi_seizure    mean 0.3188   pass   0.0%   n=8
    short_event      mean 0.2500   pass   0.0%   n=7
    single_seizure   mean 0.3911   pass  20.0%   n=5
