Yes, the symptoms you describe are consistent with two separate problems being conflated.

Diarisation quality alignment and cross-recording speaker re-identification.

I'm checking the current assembly AI and deep RAM capabilities because both have changed quite a bit,

and the right answer may be a hybrid rather than simply replacing Piano.

Yes, I think you have diagnosed the architecture problem fairly well.

I would not simply loosen the Q-RAM matching threshold dramatically as the first fix.

One match in 30 is far too conservative, but if you just lower the threshold,

you may turn mostly unidentified into the much worse failure mode of confidently attributing things to the wrong person.

There are really three stages in Briar that should be treated separately.

One, transcription. What words were spoken?

Two, diarisation. Which bits of the audio belong to the same speaker within this recording?

Three, speaker recognition re-identification.

Is this voice the moon, your father, you, etc., based on people's recordings?

Whisper deals mainly with stage one.

Fire Note is giving you stages two, and indirectly through its embeddings, the raw material for stage two.