Self-test evidence — claude-sonnet-5 via Claude Code, one run per instance
================================================================================

Configuration
  agent   : claude -p "$(cat {prompt})" --model claude-sonnet-5 --dangerously-skip-permissions
  harness : baseline/run_baseline.py, per-solver isolation, no network
  repeats : 1 per instance (a smoke test, not a benchmark measurement)

Files
  sonnet5.json              raw run output, 25 records. NOTE: `base` and
                            `eeg_case_df090a` are the SAME recording (base is the
                            published alias of df090a), so this file scores that
                            one recording twice (base 0.3663, df090a 0.2112) — a
                            good illustration of the run-to-run swing the task has.
  sonnet5_24distinct.json   the 24 distinct recordings (base alias dropped,
                            df090a kept). This is what the README section 7 and
                            the task-card summary quote.
  aggregate_summary.txt     aggregate over the 24 distinct recordings.
  evidence_logs/            two agent transcripts showing the real analysis:
                              eeg_case_e0755c (0.98) — found the single seizure,
                                rejected artifact on a broad rescan
                              base / df090a (0.37/0.21) — two seizures detected,
                                candidate windows screened and excluded as artifact

Headline (24 distinct recordings)
  subject-level mean   37.03 / 100  (sd 32.51, 15 subjects)   <- quote this
  instance-level mean  35.18 / 100  (median 18.06)
  pass rate            12.5% overall, 5.0% on positive recordings
  pooled recall        0.556 (20 of 36 events)
  pooled precision     0.500
  false positives / h  0.947
  median onset error   4.0 s
  by class: hard_negative 54.69 (50% pass) · single_seizure 39.11 (20%) ·
            multi_seizure 31.88 (0%) · short_event 25.00 (0%)
  two hardest classes, 15 instances: mean 28.67, 0 of 15 passed
