Format
paper.cssacademic print sheet · serif
The preprint series is where findings get written up so they need not be rediscovered. Each paper regenerates its tables and figures deterministically from a named, local analysis tool — and every claim is labelled simulation, real-detector, or real-data so honesty is structural, not optional.
The suite is an instrument; the papers are what it measures the world with. Where the -Dex nodes and the validation harness produce numbers, the preprint series turns a number into a documented, reproducible claim — each tied to the exact tool that generated it.
The papers live under papers/ in a deliberately different visual register from the dark suite — a white academic sheet, serif type — because their reader is a referee, not a dashboard user. They share one stylesheet and one section structure.
The pipeline keeps the claim and its evidence inseparable: the finding comes out of a live tool, the figures are exported from that same run, and the write-up follows a fixed structure that forces a stated limitation.
A result surfaces while running an analysis tool (e.g. the ODI-4/AHI calibration, the HRV confound). Nothing is written up that a tool can’t produce.
The tool exports its figures (PNG) and CSV/JSON. These are the paper’s actual evidence — reproducible from the same seed and committed corpus.
Abstract · Introduction · Methods · Results · Discussion + Limitations · Reproducibility · References — the same skeleton every paper, so a referee finds the limitation fast.
Every claim is tagged simulation / real-detector / real-data, and the limitation that bounds it is stated up front (e.g. “n=5 real nights → needs NSRR PSG to publish”).
Add the draft to papers/papers.html with its status and its tool, so the series stays a single navigable whole.
Two pilot drafts and one parked perspective. The drafts are complete on synthetic / small real pilots and name the real-data study that would let them publish.
The shipped OxyDex ODI-4 detector recovers ≈¼ of scored events (slope 0.23, R² 0.93); the AHI ≈ ODI-4 × 1.1 surrogate has LOO-RMSE 15.2/h, halved to 7.2/h by a re-fit correction. Includes a sample-size analysis: ≈150–300 paired nights to publish.
Real PulseDex rMSSD falls ≈3.8 ms/decade of age and ≈2.9 ms per 10 AHI — comparable independent weights. A single-metric rMSSD screen scores AUC 0.62 (19% old-and-healthy false flags); an age-adjustment recovers it to 0.68.
A parking note, not a result: single-channel waveform synthesis is solved elsewhere; the open frontier the harness exposes is timestamp pathology, cross-node temporal coherence, and provenance metadata. Names the narrowest first paper.
These are working drafts, not journal articles. The framing rules exist precisely so a reader is never misled about how far a claim reaches.
Drafts in a working series — not clinical guidance, not validated for diagnosis. The status pill on each says exactly how finished it is.
The findings stand on synthetic corpora and small real pilots (n=5 real nights). Each draft names the real-PSG study (NSRR/PhysioNet under DUA) that would let it publish.
Every figure regenerates deterministically from a committed corpus + a named tool. A reader can re-run and get the same numbers — the point of the series.
The index flags an OxyDex processNight hang on heavy-dropout nights, found while bootstrapping the ODI pool. Open findings are listed, not hidden.