The problem

Congressional staff routinely have to understand a long bill, amendment, or new version on a markup or floor clock. Existing redline tools can show which words moved, but they stop before answering which changes affect a particular member, committee, or district. Commercial summaries try to close that gap with generated prose, then warn users to verify the result against the bill—the exact work the summary was meant to save.

The approach

Winnow treats legislation as structured text. It computes exact diffs from official legislative XML, identifies changes such as funding amounts, deadlines, eligibility language, new authorities, and repeals, and ranks those changes at the provision level for a particular user. Claims are assembled from the bill’s own language, figures extracted by code, and fixed sentence patterns. When the text cannot support a conclusion, Winnow points to the relevant language instead of inventing one.

The same underlying diff can compare a bill with current law, an amendment with the bill it changes, or House and Senate versions of the same idea. A provision-first record could eventually follow an idea across bills and Congresses, show related sections that may be affected by a change, and alert staff when language relevant to their portfolio moves.

Where it stands

Winnow is in product development. The current design establishes the trust boundary—verify rather than generate—and the core jobs for congressional staff, advocates, and constituents. The next work is to turn that design into a reliable end-to-end system: acquiring and normalizing legislative text across both chambers, measuring the quality of provision matching and relevance ranking, and testing whether the resulting view changes what a user can understand under real time pressure.

The problem

Most political news is national even when the practical effects of policy are local. Congressional offices and constituents need a dependable account of what federal decisions mean for a particular state or district, but assembling one requires monitoring many sources, separating meaningful developments from repetition, and connecting each item to the right geography. Small offices rarely have the time or tooling to do that every day.

The approach

535-News is a location-aware briefing system for all 50 states and 435 congressional districts. Its pipeline ingests a curated roster of feeds, retains source text with rights metadata, extracts and ranks relevant material, maps stories to places, and assembles daily editions. The public product is designed around state pages, district pages, edition archives, and email subscriptions, with the source and location logic kept inspectable rather than hidden behind a generic news feed.

The same system can serve two related audiences. Congressional offices get a compact view of developments that may matter to their constituents; residents get a briefing organized around where they live rather than the incentives of a national news cycle. Administrative tools make source decisions, pipeline runs, and location corrections reviewable by a person.

Where it stands

An end-to-end prototype exists across ingestion, retrieval, generation, rendering, archives, subscriptions, and delivery operations, and it has produced sample editions for real districts and issue areas. It is not yet an operating news service and has no subscribers. Development now needs to prove the editorial standard, source coverage, geographic accuracy, and daily operating model before asking people or offices to rely on it.

The problem

The people trying to improve congressional technology work across institutional offices, civic organizations, vendors, think tanks, procurement systems, press outlets, and informal practitioner communities. Their needs, pilots, launches, and funding signals appear in different places and on different schedules. Without a shared view, builders repeat work, miss changes in demand, and struggle to distinguish a real institutional opening from general enthusiasm about civic technology.

The approach

Civic Tech Radar is a structured horizon scan for signals about Congress and the actors who interact with it. A signal must identify an actor, a stage—from stated need through build, pilot, launch, funding, or legislation—and a traceable source. The watchlist spans Hill institutions, civic groups, think tanks, vendors, practitioner communities, press, procurement, and public code repositories.

Each accepted signal receives a confidence rating, an Article One relevance rating, and a stage. Exact links and normalized titles support deduplication; a later stage creates a new historical row instead of overwriting the earlier observation. The process runs on a six-hour cadence and produces both structured records and a digest, while blocked or quiet sources are logged rather than silently disappearing.

Where it stands

The internal instrument has operated across a 31-source watchlist and recorded 143 sourced signals over 79 unattended runs as of May 2026. Its methodology, scoring, and failure handling are defined, but coverage still has blind spots: paywalled reporting, social platforms that resist reliable collection, irregular institutional publishing, and signals that require human context. Development is focused on making the radar more decision-useful for Article One and its collaborators, not merely larger.

The problem

Federal agencies regularly invite the public to comment on proposed rules, notices, and requests for information. The Federal Register makes those opportunities public, but the stream is too broad for most organizations to monitor closely. A relevant docket can pass unnoticed, and a late discovery leaves too little time to decide whether to respond, coordinate expertise, and write something useful.

The approach

The pipeline reads every open Federal Register comment opportunity, normalizes the official record, and filters it for a chosen policy domain. Deterministic code owns the canonical facts: document identifiers, agencies, source links, action links, and comment deadlines. Analytical inputs can help judge relevance or prepare a short explanation, but the feed builder rejects stale source hashes, malformed artifacts, and any attempt to replace authoritative fields with model output.

The result is a searchable, refreshable feed of opportunities with deadlines and actions attached. The architecture preserves raw pulls, normalized records, analysis inputs, validated outputs, and build reports so a user can trace a displayed opportunity back to the Federal Register rather than trusting an opaque recommendation.

Where it stands

A pilot in the food and agriculture vertical processed 1,220 Federal Register documents and surfaced 89 relevant opportunities. The deterministic pull, normalization, validation, and feed-building path is working, alongside a protected draft interface and scheduled refresh design. It is not yet a promoted production service. The next test is whether domain experts find the surfaced set complete enough, precise enough, and early enough to change whether they participate, with enough time to prepare a strong response.