Why Locus
When a goal depends on a place, the hard part is not only finding facts. It is knowing which city, county, state, GIS, permit, service-request, code, policy, transportation, environmental, legislative, or community sources matter for that exact place. Neither a person nor an agent reliably knows that on their own.
Locus exists for that gap. Its job is intent-localization: map a place-dependent goal to the relevant localized public context and return cited signals, coverage limits, caveats, and what to check next.
Why this beats the alternatives
An agent has three ways to do this itself: its own training data, generic web search and page reading, or raw open-data portals. For local civic facts all three are weak. Such facts are long-tail, they change over time, they are tied to a specific place, and the open web is noisy.
- A verified source registry, not model memory. Locus localizes from verified source cards for the jurisdiction, because long-tail recall is where models fail.
- Live fetch, not a frozen cutoff. Locus reads live official records (briefly cached) so time-sensitive answers stay current.
- Cited, with provenance. Every record carries a source name, a locator, and a timestamp, or it is dropped.
- Claims tied to records, not free invention. Facts and numbers come from the live records, not the model. The deterministic trend brief instantiates only a fixed set of allowed claims from the computed series and hard-gates the rest; the composed report and chat use a language model for the wording, but every fact must cite a returned record or it is dropped, and outputs are gated to refuse safe/unsafe, valuation, and causal verdicts. That constrains the hallucination failure mode instead of trusting the model to be right.
- Coverage before charge. Unsupported places get a free diagnostic that names the missing coverage, so a thin place becomes a source-discovery task rather than a paid dead end.
For a token-bound and cost-bound agent, the tool that returns higher quality per token and per dollar is the one a rational router converges on, and that advantage widens as questions get harder to localize and as coverage deepens.
Why it matters to people
Local context is a classic information-asymmetry problem: the landlord, the seller, and the city all know more about a place than the person considering it. A tool that maps a goal to the relevant cited records narrows that gap before a person spends hours researching, touring, underwriting, or advising. Locus surfaces the records worth checking and the questions worth asking. It does not make the decision.
The full argument, with the academic citations and the steelmanned case against Locus, lives in the repository at docs/WHY_LOCUS.md.