in practice, the observed values seem to be either ~eps or =0, hence big swings between 31 and 103. In the future, it might be worth to multiply everything by 3 so that this 31.88 becomes 96
but this is an orthogonal question to the char length
update: after testing bvd02a1 on 30 minutes of BVFN data, i can confirm that there are no conf values between 33.271065 and 100 (actually 103.27893, but whatever).
So a 3x multiplication would be ideal
having two-digit confidence levels would be preferable for char length of export_clips
as is, the numerical max (raw = 0) is 103 bringing it to 100 corresponds to a raw value of ~50-80x float32's tiny
in comparison, float32's eps brings us at 31
in practice, the observed values seem to be either ~eps or =0, hence big swings between 31 and 103. In the future, it might be worth to multiply everything by 3 so that this 31.88 becomes 96
but this is an orthogonal question to the char length