{
  "specified_utc": "2026-09-19T04:32:33.896890+00:00",
  "status": "Secondary validation added after synthetic pilot results, before any natural-window model outcomes.",
  "selection": "Enumerate all nonoverlapping3-second grid windows of the147 strongly annotated CEB soundscape test recordings. Exactly two annotated species, all events of one species before all events of the other, with >=0.1s gap. Metadata-only count was366 windows in88files/38species. Keep all eligible windows, no score-based source expansion.",
  "intervention": "Use complementary temporal masks transitioning only inside the annotated silent gap, leaving all target annotated waveform samples unchanged. Masked single-source components sum to the original natural window. Compare each target in its own masked component to the original window. Repeat with common original-window min/max normalization.",
  "report": "Report all candidate target score changes, eligibility counts, conditional loss at0.5 for targets detectable alone and not detected in the other-only component, and the stricter subset where both species satisfy these conditions. Also report gains from below0.5 to>=0.5. Group uncertainty by source file. No population prevalence extrapolation.",
  "limits": "Removal also removes background/context from the other side; attribution is to the disjoint-time acoustic context, not proven uniquely to recognized heterospecific identity. Reuses CEB recordings already used for other analyses; this is not a new held-out dataset."
}