goodearth_review_roster

io.github.lonniev/goodearth-mcp · Verify this server

Audit a roster against what this ground's record actually knows. Asked "are these the right pests to watch?", an agent will answer from its own training data — unverifiable, different per model, and confidently wrong exactly at the margins where a review matters. This answers from the record instead: the degree-day models USA-NPN publishes for these coordinates, and what iNaturalist has recorded nearby. Three findings: - **out_of_range** — listed, and this ground's record does not know it. - **absent** — the record knows it well and the roster does not list it. - **implausible** — an observation that cannot be right, with what makes it wrong. This is the one that matters: observations feed ``goodearth_calibration``, which shifts the heat and frost bias for the whole block, so a junk entry degrades every later answer rather than merely showing a wrong row. **Every reason is about the RECORD, never about the animal.** "Not recorded within 16 km" is a fact; "does not live here" is natural history, which Good Earth does not publish. Nothing is removed — the grower who genuinely saw the odd thing is precisely the case worth learning from, so findings are proposed and the patron decides.