{
  "specified_utc": "2026-09-19T04:23:23.372631+00:00",
  "scope": "Robustness plan before focal eligibility or mixture outcomes. These controls cannot rescue a failed primary source-diversity/effect gate.",
  "waveform_normalization": "Use the already validated prefix-removal intervention in normalization_control.py. All later weights/operations unchanged. Test at most300 heterospecific source pairs selected by fixed SHA256 norm:20260919:a:b, both temporal orders. Normalization extrema shared between both single-source baselines and mixture. Require original(normalized-prefix-equivalent) score error<2e-5 and assert bit-identical target samples after normalization. Report loss conditional on fixed-normalization baseline>=0.5.",
  "ambient_background": "Select three 3-second passages from distinct CEB test soundscape files with no positive-duration annotations intersecting the passage, by fixed SHA256 ranking of file/start before model scores. Continuous background RMS=0.0015, compared with bird passage RMS approximately0.015 times global clipping-safe scale. At most200 fixed-hash heterospecific source pairs, both orders. The same background samples are present in silence, singles, and mixture. Report conditional loss for baseline>=0.5 and background target score<0.5. Never select backgrounds based on the mixture effect.",
  "same_species": "Report already prescribed independent-recording same-species mixture controls.",
  "specificity": "Independent-presence counterfactual enumerates exactly four states with probability0.25 each. Report all pair-orders and baseline-specific subset; this is a computational diagnostic, not inferred wild occupancy.",
  "limits": "If ambient control loses most baseline eligibility, report that limit. No assertion of a deployable correction without false-positive validation; no new classifier or source-separation algorithm claimed."
}