The screening layer is where population intelligence meets a person. It is deliberately a conversation, not a form, twelve questions across six life domains, scored against validated instruments in real time, with the district's own FEP score used to calibrate the referral threshold.
Older adults systematically under-report difficulty, for two well-documented reasons: the fear of being a burden, and the fear of what honest disclosure might trigger. A tick-box form gives that minimisation nowhere to surface. A conversation that listens to how something is said, hesitation, qualification, the things skirted around, can detect the gap between what is reported and what is true.
This is not a soft preference. It is consistent with emerging evidence that subjective signals carry real predictive weight: a 2025 study in Geriatric Nursing testing nine machine-learning models found that self-reported experience (pain, mood, functional confidence) predicted frailty at least as well as objective physical measures. A single study does not settle the question, but it points the same way as the clinical rationale, that what a standardised form cannot elicit may be exactly what matters most.
PRISMA-7 and the FRAIL Scale are mapped across the twelve questions and scored in real time, so the person experiences a natural conversation while the system produces a defensible clinical score. The upstream district FEP score is injected as context: in higher-risk districts the referral threshold is lowered, reflecting a higher prior probability of genuine need. The output is three things from one conversation, a Wellness Guide for the person, a structured referral for the frailty team, and, with explicit consent only, an anonymised population signal.
Prototype, not validated instrument. The screen is a working prototype grounded in validated frameworks; it has not yet been formally validated against those instruments in a published community study. That study is the necessary next step.
Interpretation by a language model carries risk. Detecting minimisation is the tool's strength and its hazard: it must be tested adversarially against minimisers, cognitive impairment, hearing difficulty and carers answering over the person before it can be trusted at scale.
Threshold calibration by FEP is a design choice. Lowering the referral bar in high-FEP districts is clinically defensible but will, by construction, generate more referrals there, a feature that must be matched by downstream capacity.