Project

My Government.

Type in your ZIP code and see what every layer of your government — from your village and school board up to Washington — actually did this week, straight from their own documents.

Try it (US) Try it (Switzerland)

The idea

Government doesn't happen in one place. The same address is governed by a village or city, a school district, a county, a state, and the federal government — and almost nobody can name what any of them did last week. My Government starts from the one identifier everyone knows, a ZIP code, and assembles the whole stack: every jurisdiction that governs that address, on one page.

Each week it reads the primary record — meeting minutes and local notices at the bottom of the pyramid, the Federal Register, Congress, and executive orders at the top — and files what it finds by how final it is, so a proposal never masquerades as a law. Its scorecards track the other half of the story: what a government said it would do — strategies, budget resolutions, executive orders — and, dated and sourced, what was actually done about it since.

Deadlines to speak up — comment periods, hearings, elections — come as a calendar you can subscribe to. Every claim links to the government's own document, with the quote checked word-for-word against the source. The site states what happened and stops there; the judging is left to you.

Under the hood

My Government is one codebase running ~30 government-watching tenants — from Michigan villages and school districts to counties, states, the U.S. federal government, and a parallel Swiss deployment — each defined by a declarative spec (sources, prompts, identity) rather than code. A jurisdiction pyramid maps ZIP → place → ancestry, and a composition engine renders any address's full stack (community ⊃ county ⊃ state ⊃ federal) as one dashboard from the same per-tenant data.

Weekly pipelines pull primary sources through per-source adapters (registers, legislatures, courts, gazettes, municipal minutes), then an LLM pipeline — tiered by task and routed across providers by cost — filters for relevance, extracts structured findings ranked by legal finality, and synthesizes topic areas. Everything an LLM asserts is mechanically guarded: quotes must appear verbatim in the source text or the row is dropped.

A doctrine layer ingests stated-intent documents (strategies, budget resolutions, executive orders), extracts their commitments, and builds evidence trails in both directions — bounded, resumable, budget-capped passes link intents to the actions that carried them out, while a corpus-wide citation sweep discovers doctrine the record keeps referring to, researches and mechanically verifies each candidate (official domain, identifier-in-text), and stages it for a human curatorial confirm.

New places onboard through a probe ladder with canary-gated auto-promotion; local news sources are discovered and feed-verified against a county registry. Per-tenant SQLite, deploy-free registry overlays with human confirm gates, and public iCal/RSS participation feeds round it out. Editorial invariant throughout: state, don't grade.

Try it

Two live deployments, running the same codebase against different jurisdiction sets.

Jurisdiction Watch

The U.S. deployment — federal, states, counties, municipalities, and school districts.

Open Jurisdiction Watch

CH Gov Watch

The Swiss deployment — federal, cantonal, and municipal jurisdictions.

Open CH Gov Watch