Site-selection scoring: named partners (schools, hospitals) need distance features with their own validation, and the explanation must check existing offices
Pattern from a location-scoring agent (H3 grid, weighted percentile scores plus a Huff net-new model). (1) When the intake names specific partners (the high schools a sports-medicine practice covers, the hospital where its surgeons operate), do not send them to an LLM dataset search. Match the names deterministically against an official point layer: normalize, drop generic words like HIGH SCHOOL or H S, match the longest names first and use up the matched words so a bare district name does not also match. Then build a distance_to_nearest feature filtered to those exact names, with near counted as better. (2) A generic ingest check such as needing at least 5 points rejects a two-school partner layer. Validate named-partner layers on every named place being found plus full grid coverage instead. (3) A heavily weighted partner-proximity factor can still lose to net-new demand when an existing office already sits near the partners, because sites there would mostly cannibalize. That is correct, but a reviewer reads it as a bug unless the explanation compares the existing office distance with the candidate distance. Render that comparison from stored values.