geog.ai turns a geographic question into a capability plan, a fit-for-purpose evidence set, and a ranked answer. It does not confuse a map with a conclusion.
The planner decomposes a decision type into spatial capabilities, selects deterministic tools, and records each result. A logistics siting question and a resident-reachability question should not silently use the same evidence.
Every decision type declares which evidence is required, optional, commercial-only, or unavailable. The manifest travels with the answer so a reviewer can see the boundary of the claim.
AUTHORITATIVEOBSERVATIONALADVISORYDERIVED
These labels are not decoration. Terrain from an authoritative source is a different kind of fact than an accessibility proxy derived from a road graph. geog.ai keeps that distinction in the UI and provenance chain.
Source, vintage, provider, fitness, transforms, and the capability that consumed the evidence are preserved. A score without a chain is only an opinion with decimals.
Missing parcels, stale imagery, rejected fitness, and commercial gaps are surfaced as gaps. The engine does not paper over uncertainty with a confident sentence.
Each capability is translated here by what it does for the answer, not what it is called internally.
For each decision type, the engine picks only the data sources that are fit for purpose — not every available layer. This prevents irrelevant evidence from diluting the score.
Before a source is used, its vintage, resolution, geographic coverage, and trust level are checked against the decision's requirements. A source that fails fitness is flagged, not silently substituted.
Routing, reachability, flood exposure, seismic risk, terrain analysis, and building intensity are computed deterministically — not inferred from model embeddings.
Each candidate receives a composite score built from weighted capability outputs. The weights and the capability list are part of the evidence manifest, not hidden inside a black box.
The engine grades the overall result based on what evidence was actually available, complete, and authoritative. A strong score on weak evidence still produces SCREENING_GRADE, not DECISION_GRADE.
Missing sources, low-fitness substitutes, conflicting datasets, and commercial-only layers are all surfaced in the manifest. The reader sees the boundary of the claim alongside the claim itself.
Research and infrastructure note: Internally, these capabilities map to a Spatial Knowledge Fabric, a capability registry, deterministic adapters per evidence family, an epistemic trust taxonomy, and a synthesis layer that resolves conflicts before scoring. Those names are implementation detail; the above descriptions are the contract.