The phone is the microphone because it is the only one that does not delete
her: measured on MAY's own recordings of one room and one voice, her phone
gated 0 of 71 and 0 of 82 half-seconds, the laptop array 14-31% - and worst
exactly when she walks away, which is the audio a distance measurement needs.
What ships:
* silence is silence - the room's own floor is measured and anything near
it is reported as silence WITHOUT asking the model.
* presence threshold 0.99, not 0.8, and level normalisation to -25 dBFS.
* a verified voice (ECAPA, 0.803 on held-out text against a 0.102 null) is
the answer for its window; the scene heads are shown as what the model
said, labelled unvalidated on this channel.
* MAY's retrieval memory, its search rewritten as one matrix product -
4.43 ms a slot to 0.13 ms, verified to return the identical ranking.
* a diagnostic log per run: flat class posteriors, retrieval-vs-class-head
disagreements, device gating, room floor. Materi…
