groundtruth

Status: early. One layer of one city, and the machinery to build the rest.

The policing layer for Birmingham. Every figure is currently a hand-written placeholder pending the data import — plausible for the force named, but unverified against the sources cited, and not to be quoted. The boundary geometry is placeholder too, and says so on the map. What is real is the machinery: the verdicts, the benchmarks behind them, the estimate marked as an estimate, and the data gap that refuses to render as a pass.

I have a sneaking suspicion that there is a gap between the information available to the people making decisions about a place and the information that never reaches them. Most of it is published somewhere. It is spread across departments, released on different cycles, at incompatible geographies, in formats nobody reconciles — so decisions get made on whatever summary happened to be at hand. If that is right, policy is being set on an incomplete picture, and some of those decisions are wrong in ways nobody in the room can see.

The other half of the problem is access. Someone who lives in a city, or who simply cares about how it is run, has no single source of truth to reason from about its political, economic and social state. They are left with headlines and press releases about a place they could be reading directly.

This is an attempt at that source of truth: what is already public, assembled into one assessed and sourced view of how well a place is actually served. The gathering and the derivation are done by capable AI models working to a published protocol, and nothing they submit is published until a human has checked it.

What it is

A map-based board showing how a city is doing across the services that constitute it — starting with policing, and designed to extend to healthcare, economic activity, service delivery, electricity, water and sewage.

Each discipline defines its own dimensions: the specific questions worth asking of a place, each with a unit, the geography it must be measured at, and a direction (is higher better or worse). Every dimension is judged against published benchmarks rather than against a score invented here. Nothing is weighted or averaged into a single number, because every part of a verdict has to trace back to one value and one benchmark, each with a source.

Layer one, in progress: does a city or area have enough policing? Seven dimensions — crime outcomes, violence charge rate, recorded crime rate, estimated actual crime, emergency response time, officer availability, neighbourhood capacity.

How the data gets in

The application exposes its own MCP server. This is the collection method, not an add-on.

Any MCP-capable client — an agent from any of the major labs, or one you run yourself — can connect and:

1. get_template — read the research protocol for a layer: every dimension, its unit, the geography it must be measured at, which source tiers are acceptable, a plausible range, and the authoritative source to start from.
2. list_gaps — take the worklist: dimensions with no published value, or whose newest value has gone stale. Staleness is per dimension, set from how often that source actually publishes.
3. submit_value — submit one figure, with its provenance.
4. assess — read the verdict for a layer.

Submissions always land as pending and are never published directly. A human reviews them at /review. Rejected submissions are kept rather than deleted: what was proposed and refused is part of the record, and often the more interesting half of it.

The idea is that an application should be legible to a machine as a first-class audience, not as a scraping target. The criteria for each variable are published through the same interface that accepts the answers, so a model can find out what would count as a good figure before it goes looking for one.

What counts as a source

Provenance is not optional — the schema refuses a value without a source_url.

Sources are tiered: official_statistic, official_other, press, modelled. Tier the source honestly; press is a last resort and is treated as one. Official publications are strongly preferred, .gov.uk first where one exists, and external sources are accepted where no official figure is published — tiered accordingly.

Wikipedia is not an acceptable source. Nor is any aggregator standing between you and the figure: the source_url must resolve to the data itself, not to an article about it.

That last rule is currently policy, not code — the schema checks that a source URL exists and resolves over http(s), but no domain blocklist is enforced yet, so it rests on the review gate. Tightening it is tracked as an open issue.

A value is either measured — read directly from a source, and therefore a point — or estimated, which requires a derivation recording the method, its assumptions, its inputs and the model that produced it. An estimate without a derivation is a guess, and the database rejects it as one. These rules are enforced as constraints, not conventions.

Stack

Next.js 16, Supabase (Postgres), MapLibre, TypeScript. The MCP server is a route handler in the same app, so the interface an agent uses and the interface a person uses are backed by identical logic.

Running it

npm install
cp .env.local.example .env.local   # fill from your Supabase dashboard
npm run dev -- --port 3100

Apply supabase/migrations/ in order, then supabase/seed.sql for the policing dimensions. MCP_TOKEN is any long random string (openssl rand -hex 32); /api/mcp rejects every request without it.

Connect a Claude Code session:

claude mcp add --scope local --transport http groundtruth \
  http://localhost:3100/api/mcp --header "Authorization: Bearer $MCP_TOKEN"

Local scope, not project scope — project scope writes .mcp.json, which is committed, and the token must not land in the repo.

Honest state of things

• One layer (policing) has dimensions defined. The other layers exist as intent, not schema.
• One country's sources (UK) are wired up: data.police.uk for street-level crime, OpenStreetMap for facilities.
• Benchmarks are the weak point. A verdict is only as defensible as what it is measured against, and assembling those is the slow, unglamorous part.
• RLS is enabled with no policies: all access is server-side via the service role. Policies get written alongside the first browser-side read, not speculatively before it.
• The claude.ai web connector expects OAuth, so it cannot use the bearer token as things stand. Clients that support custom headers work today.

Licence

MIT.