Found while closing #11. tr808.md §5a recorded the reference 2–4 kHz cymbal
share as 31.3% and called our 31.3% an exact match. Both columns held our own
number. The check compared the model against itself, so it could only pass —
and it passed for a day while the voice sat 2× off, which is what #11 was
reporting at the same time.
Corrected in 5293b88: the real reference median is 63.9% (n=16).
The concern is the class, not the instance. A reference column populated
from the model is invisible — it looks like agreement, reads like a pass, and
the closer the model gets to whatever was recorded, the more convincing it
becomes. This is the same failure shape as the checkers found blind during
fault-planting: absence reading as agreement.
Worth a sweep over every dossier for figures in a reference column that no
longer trace to a source. What to look for:
- a reference figure identical to the model's measured value, especially to
the same decimal place — 31.3 vs 31.3 is what gave this one away
- a reference figure with no
n=, no IQR, and no named source pack
- a figure described as "exact" or "matched" rather than given a tolerance
The mechanical part is cheap: for each dossier row with a stated reference,
re-derive it from the pack and diff. Where a row cannot be re-derived because
no capture backs it, that is itself the finding — the row is an assertion, not
a measurement, and should say so.
Nothing here is urgent or sound-affecting; it is a documentation-integrity
pass. But it should happen before any §5 criteria are adopted from these
columns, since #11's proposed band-share criterion would have been adopted at
the wrong figure.
Found while closing #11.
tr808.md§5a recorded the reference 2–4 kHz cymbalshare as 31.3% and called our 31.3% an exact match. Both columns held our own
number. The check compared the model against itself, so it could only pass —
and it passed for a day while the voice sat 2× off, which is what #11 was
reporting at the same time.
Corrected in 5293b88: the real reference median is 63.9% (n=16).
The concern is the class, not the instance. A reference column populated
from the model is invisible — it looks like agreement, reads like a pass, and
the closer the model gets to whatever was recorded, the more convincing it
becomes. This is the same failure shape as the checkers found blind during
fault-planting: absence reading as agreement.
Worth a sweep over every dossier for figures in a reference column that no
longer trace to a source. What to look for:
the same decimal place — 31.3 vs 31.3 is what gave this one away
n=, no IQR, and no named source packThe mechanical part is cheap: for each dossier row with a stated reference,
re-derive it from the pack and diff. Where a row cannot be re-derived because
no capture backs it, that is itself the finding — the row is an assertion, not
a measurement, and should say so.
Nothing here is urgent or sound-affecting; it is a documentation-integrity
pass. But it should happen before any §5 criteria are adopted from these
columns, since #11's proposed band-share criterion would have been adopted at
the wrong figure.