From Clinical Recommendation to Behavioural Action: How a Consent Handshake Determines Which Patients Reach Coaching

A clinical decision support system can tell a patient to change a behaviour, but that only helps if the patient agrees to be coached and the software correctly recognises the agreement. In a conversational coaching agent, recognition works as a handshake: the coach proposes a goal, the patient replies, and a classifier decides whether the reply means yes, no, later or stop. Every path into coaching runs through that decision, which makes it an eligibility criterion. It appears in no protocol, and it cannot be found by reading the code, because it exists only when the controller and the classifier run together. We compute it. The controller has a finite number of internal states, so we can list all of them and turn ``will this patient be coached?'' into an exact calculation on a Markov chain. The calculation needs one measured input: how the classifier really labels replies. It returns two quantities, both available before any patient is enrolled. ${Admission}$ is the chance of reaching coaching within ten patient turns. ${Dose}$ is how many coaching turns a patient can expect in a twenty-turn session. Both depend on the software as much as on the patient. With the model unavailable, admission ranges from 0.065 to 0.842, and dose varies more than nine-fold, between two constructed patients who both agree and differ only in how tentatively they say so. When a language model is available, admission rises by 0.924 for the most tentative patient and 0.115 for the most direct. So availability does not lift everyone equally; it changes who is served worst. Across the eight patients the change runs from $-0.079$ to $+0.924$, and only three of the eight can be told apart from zero. We price three repairs by undoing each one and recomputing. Repairing the classifier makes a separate structural flaw ten times ${cheaper}$ rather than more costly, because that flaw sits on the path a patient takes after a pause. The method also bounds itself: the classifier that admits the most patients is one that mistakes hesitation for consent, and no admission measure can detect this, because that mistake improves all of them. The calculation is exact once the labels are fixed, but the labels come from 90 patient turns, and resampling those turns (a bootstrap) leaves zero inside the range for every repair price. Every patient is constructed, the controller is in development, and we do not claim that coaching helps once delivered.