Pre-registration v1.0Not yet locked25 participants16 weekly occasions

Within-person measurement thresholds: a prospective validation.

Whether thresholds calibrated to an individual detect real change more accurately than a fixed published cut-off, tested against thresholds committed before any data is collected.

Hypotheses

Three, each with a pre-specified threshold. H0 is a real possibility.

H1
Primary
Conformal intervals from a participant's own history achieve realised coverage within ±0.03 of the stated 0.90.
H2
Comparison
Within-person thresholds outperform the published fixed MDC of 5.0 kg on the same observations, measured by F1.
H3
False alarms
Participants whose measurements are genuinely stable are flagged as changed at a rate not exceeding 0.15.

Locked thresholds8

Hashed at lock time. Any later change breaks the hash and voids every result computed against it.

Primary measurand: grip_kg — Published MDC 5.00 kg; weekly, unsupervised, no clinic
conformal_alpha: 0.10 — 90% stated coverage
mdc (fixed comparator): 5.00 — The published figure H2 tests against
min_occasions: 9 — ⌈1/α⌉−1; fewer cannot certify 90% coverage
trend_alpha: 0.10 — Trajectory verdicts
window: 7 — Occasions in the comparison window
meaningful_threshold: 2.00 — Smallest change a participant would act on
Holdout fraction: 0.20 — SHA-256 of (seed, occasion); not re-drawable

Per-participant thresholds are derived from their own baseline scatter during weeks 3–4 and locked at the same moment, before any verdict is shown to anyone.

Baseline verdicts are withheld

Weeks 3–4 collect measurements and show nothing.

Why

A threshold learned from influenced data is not a threshold.

If a participant is told what the system thinks before their baseline is set, the threshold that follows is estimated from a sample the system shaped. Nothing in a conventional analysis would reveal it.

Schedule

Two weeks of setup, two of silence, sixteen of measurement.

Weeks 1–2 enrolment and instrument training. Weeks 3–4 baseline, no verdicts. Weeks 5–20 the study proper. Weeks 21–22 analysis against the locked registration.

Outcomes5

All five reported regardless of direction. Selective reporting voids the registration.

01
Realised conformal coverage
Against the stated 0.90, tolerance ±0.03. The primary endpoint; everything else rests on it.
02
False-flag rate
Among participants with no detectable trend.
03
Head-to-head against the fixed 5.0 kg cut-off
Scored by F1, never PPV alone. A near-silent rule scores high PPV by default: in simulation a fixed cut-off reached 100% PPV while catching 3 of 69 real changes.
04
Actionability
Proportion of verdicts that changed what a participant did, self-reported at the following occasion.
05
Holdout performance
MAE and coverage on the 20% of occasions no estimator ever saw.

Stopping rules

Decided before there is a reason to argue about them.

139
Interim gate
If realised coverage falls outside 0.85–0.95 at 139 occasions, the study stops and the cause is diagnosed. Continuing to 385 to obtain a tighter estimate of a broken number wastes 25 people's time.
F1
Futility
If the fixed cut-off leads on F1 at the interim gate and the comparison is not degenerate, the study completes but H2 is reported as rejected. Personalisation not earning its complexity is a finding.
SHA
Integrity
If the registration hash fails verification, every result computed against it is void and the study restarts under a new registration.

What this study cannot establish

Stated so nobody has to infer it from silence.

—
Nothing about genotype-predicted response
A randomised trial already found the null (β = 0.06, 95% CI −1.33 to 1.44, P = 0.94). No genomic endpoint appears here.
—
No clinical or diagnostic claim
The output is change from a participant's own baseline with a stated false-alarm rate.
—
No generalisation beyond the listed measurands
Coverage on one measurand in a self-selected group of 25 is not coverage in general.
—
No efficacy claim for any intervention
Participants are not assigned to interventions.