Where is a pilot worth testing?
Before a plan spends anything, Population Fit scores its population across six dimensions and returns one of three honest answers: pilot candidate, discovery required, or not a current fit, with the evidence behind each.
Today a rules-based template writes the recommendation. The Claude connection is built but switched off, and the evaluation harness has to pass before it can be turned on.
- ScoringWorkingSix-dimension fit score + decision rules. Need alone can never produce a pilot; feasibility, existing programs, and missing data all move the answer.
- EvidenceWorkingEvidence packet with stable IDs. Supporting, counter, and missing facts, each traceable to a displayed signal.
- AI seamWorking · templateThe recommendation is written from the packet. A template writes it today; a Claude adapter can take over. Neither can change the verdict or cite evidence that isn’t in the packet.
- GuardrailWorkingRequest-scope classifier. Refuses individual targeting and clinical or eligibility use, and offers a population-level alternative.
- EvalWorking20-case golden suite + release gate. Safety gates must hit 100%, and injected model faults flip the gate to Not ready.