Wearable signal
LIVEA blood analyser fires hundreds of thousands of cells through a laser and reports twelve numbers. A wearable does the same thing to movement: it measures a rich stream and reports a summary. This is what the discarded part carries, measured with the same instrument that scores the medical benchmarks — literally the same metric, not an analogy. Opens standalone.
The same metric, on a different sensor
Logistic regression on the discarded 3-8 Hz bandThe three placements / four sample rates are the SAME subject at the SAME instant, which is the spec's Protocol A design (paired re-acquisition). It is scored as Protocol B -- the weaker, public-data protocol -- so the number is directly comparable with the imaging records and never overstated. Because the settings are paired, every cell is also leave-one-subject-out: the model is fit on the source setting using nine patients and tested on the target setting for the tenth. Without that, a paired transfer matrix would measure subject memorisation.
Why this is the same instrument and not an analogy
Dimension 2 asks whether a model survives being re-acquired somewhere else. On imaging that means a different plane of the same volumes. On a wearable it means a different place on the same body, at the same instant. The definition does not change, and neither does the code path: the implementation that produced the sensor record above recomputes mean retention, worst retention, the score and the chance correction for every published cross-site record from its own formulas, and reproduces all fifteen to 1e-9 — seven imaging, eight tabular, across three site families.
That is checkable rather than assertable. From plan_a/:
Exit 0 only if every published cross-site record reproduces and there was at least one to reproduce. This check has now been wrong twice, and both are worth stating. Until 29 Jul 2026 it printed “0 of 0 … reproduced” and still exited 0, because the 28 Jul repository reorganisation left it globbing a path that no longer existed. Until 31 Jul it printed “7 of 15 … reproduced” and still exited 0: main() returned the exit code and the entrypoint discarded it, and the chance correction was hardcoded to the 11 organ-CT classes, so the eight 2-class tabular records added since were measured against a constant that was never about them. The records were correct; the check was not. It now takes each site’s declared class count from the task registry that produced the records, fails on any site it cannot resolve one for, and returns its exit code. A check whose failure is invisible is worse than no check.
Then it caught us
The direction only being useful is not the interesting part. A caveat MedEval-1 wrote for scanners — that a retention number is meaningless once the target acquisition drops below what the signal needs — applied unchanged to sample rate, and it caught a real failure in the sensor result: the band survives down to 16 Hz and collapses at 8 Hz, below its own Nyquist limit. That collapse is published on the page below rather than smoothed, and it was found by a rule written for a different modality entirely. That is what it means for these to be one platform, and it is the reason the claim is worth making at all. See the same dimension on the medical suite →