Free Agentic Publication Digester (FAPD)

Ask DeepWiki (opens in a new tab)

FAPD is an automated pipeline that reads the official publications of the United States federal government — congressional floor proceedings, bills, the Federal Register, enacted laws, federal court opinions, and agency press releases — and produces a daily, cited, opinion-agnostic digest for two readerships at once: people, and AI agents researching what the federal government actually did on a given day.

Every digest covers one publication day across all three branches. Every item carries a citation to the official record and names the mechanical rule that selected it. Everything not summarized is counted. Nothing is silently omitted.

Status (2026-07-30)

The authoritative numbers block — where any other figure in the repository disagrees, this dated snapshot is the current one.

Live site: https://fapd.info — served from a Docker stack on a VPS; GitHub holds the repository, CI, and the integrity record.
Source registry: 127 sources — 30 active (14 web feeds and newsrooms, 11 email bulletins, 5 govinfo collections), 76 planned, 19 recorded unavailable, 2 evaluated and excluded.
Latest digest: 2026-07-29.
Test suite: 377 tests (verified by collection 2026-07-30).

Philosophy

Built exclusively on the official record

Everything here derives from official government publications — the record a government produces precisely in order to make it public. Congress prints its proceedings; agencies publish their rules and announcements; courts post their opinions. FAPD uncovers nothing: it takes what is already published and makes it easier to find, read, verify, and — for AI agents — ingest. Primary sources are the ground truth. News coverage and commentary are never ingested, never consulted, never cited.

This is why the project is legitimate infrastructure rather than scraping-as-adversarial-sport: the pipeline consumes only the channels governments built for the purpose of being read, identifies itself honestly on every request, and treats a refusal as an answer. A government that publishes its actions has already voted for this kind of reader to exist.

Two commitments, held in tension

Digestible — one person should be able to absorb the day's significant official activity in one sitting.
Faithful — no summary without a citation; nothing significant silently dropped; no editorial spin, ever.

Summarization is where bias creeps in, so the tension is resolved by machinery, not judgment calls:

Mechanical, party-blind selection. What appears in a digest is decided by coded, versioned rules — floor time consumed, recorded votes taken, regulatory economic-significance designations, stage of process — never by subject-matter preference. Every listed item states its rule inline ("Included because: …").
A citation on every item, linking to the official govinfo package or the originating agency document.
A mandatory Coverage Statement closing every digest: what was published that day, what was summarized, what was deliberately not, with counts that must reconcile arithmetically against the database.
Labeled model layers. All model-generated text is labeled in place and linted against a banned lexicon of loaded adjectives and motive attribution. Official text quoted verbatim is identified as such.

A digest that fails any validation check — a citation that doesn't resolve, coverage arithmetic that doesn't reconcile, a banned term in our prose, a missing inclusion rule — is not published. There is no override.

A publishing house for humans and agents

AI agents answering "what did the federal government do on date D?" today either crawl official sites themselves (multiplying load on public infrastructure) or lean on secondhand coverage. FAPD offers a third path: one disciplined, identified, budgeted crawler reads the record once, and agents ingest the summarized, cited, coverage-accounted result — stable URLs, llms.txt, a machine-readable digest index, an Atom feed, and provenance manifests with SHA-256 records for verifying captured content.

One ask travels with the data: for claims, cite the underlying official source (each item carries its govinfo ID or agency URL); cite this project for the aggregation. The digest is a route to the record, never a replacement for it.

Forkable by design

The editorial gates, provenance model, access discipline, and adapter seam are jurisdiction-neutral. Pointing the codebase at another government's official publication interfaces means replacing the source registry and the parsers — not the rules. See docs/adding-sources.md.

How the digester obtains content

The acquisition layer is a set of automated agents with one governing principle: we are guests on public infrastructure, and every rule below is enforced in code, not by operator discipline.

The source registry is the scope authority

sources/registry.yaml (rendered as SOURCES.md) records every federal source ingested, planned, evaluated-and-excluded, or found unavailable — currently 127 sources across a tiered universe, so "how comprehensive is coverage?" is a measurement, not a claim. Honest statuses matter: a source that blocks our honestly-identified client is recorded unavailable with the observed behavior. That fact is itself published accountability data.

The closed third — and the standing effort to open it. A measured share of the federal source universe currently refuses honestly-identified automated access (the July 2026 probe found 22 of 72 non-govinfo sources closed behind WAFs or robots disallows). We treat that as the project's ongoing engagement agenda, not its boundary: publishers' own access documentation keeps revealing doors on other hosts and paths (that alone re-opened FCC, Commerce, and NOAA candidates); verdicts are re-probed as sites change; and we engage agency web and API teams directly to advocate for safe, sane automated access to what they already publish for the public. Coverage grows by doors opening — never by evasion.

That effort produced its first result in July 2026: 11 agencies whose web channels refuse us now have a working input path through their own email bulletins — Treasury, USDA, EPA, SSA, DOT, FAA, NHTSA, DEA, ATF, the Coast Guard, and HUD's Inspector General. The blocked web entries stay in the registry exactly as they were; the email entries sit beside them as siblings. A refusal recorded is never quietly erased by a success elsewhere.

Five onboarding gates

No source is ingested on a hunch. Each one is (1) registered with identity, tier, and URLs; (2) probed end-to-end through the identified client — robots verdict, fetch with provenance capture, feed detection, item inventory, sample extraction; (3) content-evaluated — what does this source publish in total, and what fraction will ingestion see? Under-coverage is disclosed at onboarding, not discovered later; (4) activated into ingestion and coverage accounting; (5) re-evaluated on failures or redesigns. Documentation research precedes probing: the publisher's own developer pages, API docs, and feed directories say which door they built — and reading the sign on the front door has repeatedly beaten guessing (documented feeds that HTML autodiscovery misses; APIs on hosts a newsroom WAF never touches).

Source adapters and the access hierarchy

Real publication interfaces are irregular — feeds without stable IDs, article pages that challenge sustained automated access, content behind script-only redirects. That irregularity is absorbed at one seam: a source adapter owns exactly four decisions (what makes two sightings the same document; whether to fetch full articles or feed metadata only; how served bytes become text; what to store when no article is available), while the shared loop owns everything that must never vary — conditional requests, robots enforcement, budgets, capture, storage.

Adapters reach for access in a fixed order:

Directed programmatic access — the API, bulk-data service, or feed the agency itself publishes for machines. Always preferred: it is the channel the publisher built for exactly this.
Basic web access — the same HTML a citizen reads, through the robots-enforcing client, only where no directed channel exists.
Subscription channels the publisher pushes — agency email bulletins (GovDelivery, listservs), where a project mailbox subscribes through the agency's own signup form like any citizen. This is consent at its clearest: the publisher transmits each item to us. It is not a way around a web refusal — receiving a bulletin is never treated as permission to crawl a site that turned us away, and the refusal stays on the record. See docs/email-sources.md.
Never impersonation. No browser user-agent spoofing, no script execution, no WAF or bot-check circumvention — ever. Parsing structured data a server chose to send us is legitimate; pretending to be something we are not is not. When justice.gov answered sustained article fetching with bot-challenge interstitials, the response was to stop fetching articles and disclose the degraded mode per item — not to solve the challenge.

Politeness is a promise made to each server individually

Paced: at most 1 request/second sustained per host — and a host's robots.txt crawl-delay overrides that downward. When gao.gov asked for one request per 420 seconds, it got exactly that; the answer to the resulting slowness was fewer requests (its feed's rich official summaries), never faster ones.
Concurrent across hosts only: ingestion polls distinct hosts in parallel, each worker with its own pacing clock, so no server waits behind another's promises — and no server ever sees more than it would if it were our only source.
Continuous, with backpressure: in production a collector supervisor (src/fapd/collect.py) polls each source class on its own clock through the day — govinfo about every 30 minutes, agency feeds about every 60, the project mailbox about every 15 — always as watermark deltas with conditional requests. Past 70% of a class's daily budget its interval doubles for the rest of the day, reserving headroom for the end-of-day finalizer.
Budgeted: hard daily request caps per source class, enforced by the client itself, which refuses to exceed them mid-run.
Logged, twice: every outbound request lands in a queryable fetch log written by the HTTP client (nothing can bypass it) and in a human-readable daily access narrative. scripts/audit.py reports the footprint at any time.
Identified: a descriptive User-Agent with a contact address, on every request, to every server.

Provenance for a record that can change

GPO's record is stable; agency web content can be edited or removed without notice. So every capture is preserved under a two-hash strategy — content_sha256 over the exact served bytes (the evidentiary hash, stored content-addressed) and text_sha256 over normalized text (the change signal) — and every fetch attempt, including errors and robots refusals, is exported to a daily manifest committed to this repository. Each manifest chains to the previous day's by hash, so deletion or reordering of days is detectable from the files alone. New captures are additionally submitted to the Internet Archive's Wayback Machine as an independent second witness, within a budget, best-effort. A source's claimed publication date and the time we first observed it are always stored separately: the claimed date is the agency's assertion, our observation is the audit trail. What the hashes prove — and what they don't — is stated in full in PROVENANCE.md.

Inference: exactly where models are used, and where they are not

Mechanical work is code. Selection, counts, stage groupings, the entire Coverage Statement — zero model involvement, reproducible from versioned rules. An LLM call that could have been a SQL query is treated as a bug.

Official summaries come first. Federal Register agency abstracts, the Congressional Record's own Daily Digest, official bill titles and stages, court syllabi — used verbatim, identified as official text, at zero inference cost. Models write only what the record does not already summarize.

Four model layers, independently versioned (iterating on one never regenerates another):

Item summaries — for selected documents lacking an official summary; generated from extracted source text, stored durably, always cited.
Plain-speak restatements — one "In plain terms" sentence per item, derived only from the stored summary sitting directly above it, so a reader can check the interpretation against the official text in place. It adds no facts and no significance judgments. It is labeled interpretation, and an item whose restatement fails validation simply renders without one.
Day and section syntheses — the "Day in Review" and per-section quick-reads, composed from stored summaries and mechanical counts, under the strictest scrutiny.
Section tags — each digest section's Tags: line: mechanical branch and agency tags first (zero tokens), then up to three model-generated one-to-three-word discovery keys per section, generated in one batched cheap-tier call per day from the stored synopses, labeled model-derived, and linted by the same banned-lexicon gate because they render in the digest.

The banned-lexicon gate scans all generated prose against a coded list of loaded adjectives ("landmark", "controversial", "sweeping"…) and motive attribution ("in an attempt to…"). Verbatim official text and citation URLs are masked first — the gate polices our language, not the government's. A match blocks publication.

In transformation, models are the secondary tool. Turning source data into pipeline records is deterministic first — feed fields, embedded structured data, official metadata. Model inference is reserved for what programmatic shaping genuinely cannot recover, and is then budgeted, ledgered per token, prompt-versioned, and marked model-derived in metadata — never laundered into fields that read as source-provided.

The digest

One Markdown document per publication day (the canonical artifact; the static site is a derived presentation): a clickable table of contents, the Day in Review, numbered sections for floor activity, legislation, the Federal Register, enacted laws, judicial activity, and agency announcements (listed only when the agency itself dates the release on the digest day — backfill is disclosed and counted, never passed off as news), per-item citations and inclusion rules, plain-speak lines under their official counterparts, per-section tag lines, a glossary, the Coverage Statement, and a methodology footer. Sample: digests/2026-07-28.md.

On the site, each digest renders as a collapsed, plain-speak-first view: section cards whose headers carry the title, tag chips, and the plain-language synopsis, expanding to the full record on demand — still static HTML, no scripts. (The derived live page carries one inline script that renders timestamps in the reader's local time alongside UTC; it loads no external resource and stores nothing.)

Setup

sh uv sync # install dependencies into .venv cp .env.example .env # then paste your api.data.gov key uv run python scripts/verify_key.py # one-request sanity check

Get a free API key at https://api.data.gov/signup/ (emailed instantly).

Daily operation

In production (the VPS Docker stack in deploy/vps/), a collector supervisor polls sources continuously through the day and run_pipeline.py runs as the end-of-day finalizer that freezes and commits the canonical digest. Each stage also runs standalone:

sh uv run python scripts/collect.py # collector supervisor (continuous ingestion; --once for one cycle) uv run python scripts/run_pipeline.py # EOD finalizer: sync → ingest → extract → analyze → digest uv run python scripts/sync.py # govinfo delta sync only uv run python scripts/ingest_agencies.py --verbose # agency newsrooms only uv run python scripts/digest.py --date YYYY-MM-DD # (re)render one digest uv run python scripts/build_site.py # digests/*.md -> site/ (static HTML) uv run python scripts/audit.py # our server footprint, from the fetch log

Sync is watermark-based (only changes since the last run are listed) and rate-limited per GUIDE.md §4. A first run with no watermark is date-bounded; deeper history comes from govinfo's bulk-data service, never the API.

Layout

src/fapd/ pipeline code (fetch → extract → analyze → report → publish) sources/ registry.yaml — the source universe and its statuses scripts/ operational entry points provenance/ committed daily manifests (hash-chained) digests/ generated daily digests (committed, canonical) site/ derived static site (human pages + agent surfaces) deploy/vps/ the Docker stack serving fapd.info (source of truth) docs/ schema, research reports, dev notes, how to add sources data/ raw archive + SQLite (local, git-ignored) tests/

How this project is built

FAPD is developed with generative AI — Claude agents writing code, running documentation-first source research, and drafting governing documents under the operator's direction, with every commit carrying a co-author trailer. A project that labels machine-generated prose in its digests does not hide its own machine authorship; the full statement, including what AI assistance changed about where design attention went, is at docs/site/ai-development.md (published on the site as "How AI Built This").

Licensing

Code: Apache-2.0 (with NOTICE).
Content — digests, site pages, explanatory documents: CC BY 4.0. Reuse freely with credit to "FAPD — Free Agentic Publication Digester"; the attribution rule mirrors the project's own citation ethic (cite the official source for claims, FAPD for the aggregation).
Quoted official government text is public domain (17 U.S.C. § 105) — it was never ours to license.

The documents that govern this project

GUIDE.md — the governing document: mission, editorial principles, source classes, respectful-access policy, token economics, provenance threat model, roadmap. Changes to it precede implementation.
SOURCES.md — the living source guide (generated).
PROVENANCE.md — what our capture hashes prove, and honestly, what they do not.
docs/adding-sources.md — onboarding a new source, writing an adapter, or pointing FAPD at another government.
docs/devnotes/ — narrative engineering notes: how the digester came to work the way it does.
WORKLOG.md — the timestamped record of every work session. (Entries before 2026-07-28 use the project's working title, "Information Intelligence" — preserved, not rewritten.)