Specified before it is validated: a voice triage instrument, its self-invalidation rule, and the protocol that would test it

This record publishes a clinical triage instrument in full -- its inputs, its decision rules, its colour bands and its referral routing -- and publishes no evidence that it works. There is no sensitivity, no specificity, no positive predictive value and no AUC anywhere in it, because the paired data needed to compute them have not been collected. That is deliberate, and it is the point. WHAT WAS WITHDRAWN. An earlier version reduced several acoustic quantities to a single composite index scaled by a physical constant. It has been withdrawn, and not replaced by a better index. The magnitude of such a number follows from the structure of the formula rather than from the severity of what is measured, and it cannot be calibrated against any independent scale. What replaces it is a different kind of object: a deterministic decision procedure whose rules are stated in advance and whose performance is a matter for measurement. TWO COMPONENTS, KEPT APART. The estimator maps voice to a per-system score and requires validation against measured blood values; it is not validated and is not claimed here. The interpreter maps scores and a voice-derived respiratory capacity onto one of twelve patterns, a colour band and a referral; it is deterministic and requires specification rather than validation. Publishing the second without the first is possible, useful and honest. THE SELF-INVALIDATION RULE. The instrument's first rule invalidates its own output. Respiratory capacity is estimated twice -- from a sustained vowel and from a counting task -- and if the two disagree while voice quality is deviant, the respiratory estimate is declared unreliable and the shortfall attributed to the larynx rather than the lung. A person whose vocal folds close incompletely cannot sustain a vowel, and a lung capacity inferred from how long they can will read low for a reason that has nothing to do with their lungs. An instrument that knows when its own answer does not apply is the clinical form of a measurement gate. WHY PREVALENCE MUST BE STATED. At 1 % prevalence, a flag with sensitivity 0.90 and specificity 0.95 -- figures anyone would call good -- yields a positive predictive value of 0.154: six and a half people flagged per true case found. To reach a predictive value of even 0.20 at that prevalence the specificity must be 0.964. A colour band is a queue position, not a diagnosis, and this is arithmetic rather than modesty. THE PROTOCOL, PRE-REGISTERED. The primary outcome, the three null models the instrument must beat -- age and sex, the ambient noise level of the recording, and a classifier trained to predict the recording session -- the sample size, the analysis and the conditions under which the instrument would be abandoned are stated before the data exist. Pinning sensitivity to +/- 0.05 needs 139 true cases: about 2,766 participants at 5 % prevalence, and 13,830 at 1 %. THE RULE SET is deposited with exact thresholds, and writing it in a fixed notation surfaced four defects that were invisible in prose: a condition in the first rule that can never be satisfied, a rule set that is not exhaustive, three conditions that admit more than one reading, and an unspecified z-score baseline. They are reported rather than repaired, because each repair is a clinical decision. Anyone citing this record as evidence of performance is citing it wrongly. A clinician who thinks one of the twelve rules is wrong can say so today rather than after a trial, which is the whole reason for publishing it in this form.