An audited, citation-backed evidence base for every U.S. state, DC, and the five territories, so the question “which jurisdictions, and on what footing” rests on facts, not assertion.
A coordination hub has to know the ground it stands on. The landscape analysis is the foundation that lets jurisdiction selection, and, later, the live hub itself, rest on defensible, auditable facts. It is deliberately neutral: it produces the evidence and leaves the decision to a transparent step that anyone can inspect.
The same three stages run for all 56, FIPS-keyed end to end.
Deterministic Python pulls every federal source, vintage-stamps it, and computes it into eight dimensions across 37 sub-indicators. Reproducible: same inputs, same numbers.
→ partial_profiles/<fips>.partial.json
Five independent LLM agents read each jurisdiction, run on two providers so no single model's claim stands alone.
→ profiles/<fips>/<provider>/<agent>.json
Fuses quantitative + qualitative under explicit precedence rules, derives the confidence flags, and validates every profile against a fixed schema. Owns no original data; it only integrates.
→ profiles/<fips>.json
Eight dimensions, partner candidates, sector priorities, regulatory backdrop, federal hubs, access & reach geography - and a confidence ledger of every gap, low-confidence value, and stale vintage.
→ hands off to the Keystone hub
the evidence base behind jurisdiction selection (a weighted, sensitivity-tested decision memo), and the seed for the Registry, whose seats the profiles already name.
Unedited values from the actual run. Note the partner seats are the same agnostic roles the hub's Registry runs on; the analysis already names who could fill them , and that Montana’s S&P rating is honestly marked NR, never guessed.
recommended sectors
nonprofit
state agency
0 missing · 3 low-confidence · 6 stale · every value cited
recommended sectors
nonprofit
state agency
0 missing · 3 low-confidence · 7 stale · every value cited
recommended sectors
nonprofit
tribal-nations liaison
1 missing · 4 low-confidence · 5 stale · every value cited
Snapshot from the 2026-04-28 run · the live evidence base sits behind the backend's /profiles + /selection APIs.
37 sub-indicators, each tracing back to a cited federal source.
BLS QCEW · BEA SAGDP
QCEW + GVA vs AI-exposed sectors
Carnegie R1s · NSF HERD $ · land-grant
Census BTOS + maturity scout
WIOA · RAPIDS · Perkins · CC
in-jurisdiction federal hubs
CDC SVI · ACS · BIA · MSIs · geography
S&P · rainy-day · governor turnover
Each measurement carries its source, table, and retrieval date, plus a confidence grade: high, med, low, or missing.
Unknowns are marked missing, never guessed. A territory a source doesn't cover stays explicitly blank.
One FIPS-keyed schema for all 56. A state, DC, or a territory is just data, never special-cased.
It refuses to rank. It emits evidence and routing flags; the actual selection is a separate, transparent scoring step.
The landscape is the “know the terrain” layer; Keystone is the “coordinate on it” layer, and the same coordination seats the hub runs on are exactly what the analysis already names, jurisdiction by jurisdiction.