The Disclosure Sweep Agent
Agents

The Disclosure Sweep Agent

Skip to main content
Table of Contents
< All Topics
Print

The Disclosure Sweep Agent

Last updated: November 20, 2026

Federal law requires much money and influence to be disclosed: lobbying, foreign-agent registrations, campaign contributions, federal contracts, company filings and officials’ financial statements. The records are public and scattered. The disclosure sweep agent gathers them for one subject, dates and sources each row, and flags patterns that deserve a human look. It does not decide what any of it means, and it never accuses.

What it is

The disclosure sweep agent (disclosure-sweep-agent) is an AI agent with a written job description. It mirrors the investigator or data researcher: the person who pulls every disclosure on a subject so a reporter can decide what is a story.

When it runs

A human dispatches it on demand, one subject per run. The subject is a company, an official or a member of Congress. It does not run on a schedule.

What it does

  1. Classifies the subject. It decides whether the subject is a public official in that capacity, a company, or anyone else. Anyone else goes to the inclusion gate first (see below). An empty subject, or a name it cannot resolve to one entity, is a stop, not a guess.
  2. Sets a budget. Every source is queried through the research gateway, scripts/ca_research.py, under a per-run budget.
  3. Sweeps one source at a time and saves each response in its run directory:
Source What it covers
Lobbying disclosures Registrations and quarterly reports that name the subject as client or registrant
Foreign-agent registrations (FARA) Off by default; it runs only if an editor has already enabled it. Otherwise the dossier says the source is disabled
Federal Election Commission Candidate and committee totals, plus the filing schedules the client supports
Federal awards (USAspending) Prime awards to the subject; agency-level budget figures only for agencies, never for a recipient
Federal contractor registration and exclusions (SAM) Registration and exclusion records; needs a key, and if none is available the dossier says so
SEC filings (EDGAR) Filings by form type, with the form recorded on every row
House financial disclosures Members and House candidates
Executive-branch disclosures (OGE) Financial disclosures and periodic transaction reports for appointees
Senate financial disclosures Off by default, like FARA
Government audit reports and inspector-general reports Link-only: the agent lists the search addresses for a human and does not fetch them
  1. Writes the dossier. One table per source with columns for record, date, amount and URL. Every row traces to a response saved in the run directory. A source that returned nothing gets its own table with one row saying why: no records, disabled, client not registered, key missing or rate-limited.
  2. Writes the anomalies table. Patterns worth a human look, with columns for the pattern, the dossier rows it rests on, what the record shows and an innocent explanation. See the rules below.
  3. Lists links for a human to the audit and inspector-general searches it did not fetch.

Budget discipline

The federal contractor registry has a very small daily budget shared by every tool we run, so the agent makes at most one call per source per subject. A rate limit, refused request or unavailable source is recorded in the dossier as that source’s outcome and the sweep moves on. It never retries in a loop and never swaps in another source to fill a gap.

Skills it applies

Published skills:

One internal working skill, described in words only: a method for investigating federal spending and contracts.

If a required skill is missing, the agent stops and reports it; it does not run a method from memory.

Rules and gates it obeys

The inclusion gate runs first for any subject who is not a public official or a company. A fail or an unclear result means no dossier for that subject. The agent records the name and reason and moves on. The gate is not skipped for time. See Inclusion and Fairness Gates.

Wording is about sequence only. The anomalies table may say “timing coincided”, “followed within N days” or “the filing postdates the award”. It must never call a row insider trading, a quid pro quo or corrupt, and must never imply intent. Every row carries at least one plausible innocent explanation. A row the agent cannot write one for is not yet an anomaly, and the agent says what record would settle it. These match the language rules in Evidence, Sourcing and Language Rules.

Its runtime limits are in AI and Research Runtime Guardrails. The agent stops at the first blocker: a failed or unclear inclusion-gate result, a missing skill, an unregistered data client for a source the brief requires, a missing key for a required source (recorded by name only in the agent’s results), or an empty subject.

What it may never do

  • Email, call, message or otherwise contact a subject, an office, a company or a filer.
  • File anything, with anyone: no records request, complaint or tip.
  • Write into the knowledgebase outside _drafts/, or write a knowledgebase document at all unless the brief explicitly asks for one.
  • Set, export or change a switch that turns on a data source that is off by default.
  • Read the file that holds our credentials, print a key or copy one anywhere.
  • Make a network request except through scripts/ca_research.py.
  • Commit, merge, push or publish anything.

Where a human decides

The reporter decides what is a story. The dossier is a gathered, dated record and the anomalies table is a list of questions with innocent explanations attached. Neither is a finding. Turning a pattern into a published claim is an editor’s job and goes through the checks in Draft Checks and Human Review. The agent’s results file lists the subjects swept, every gate fail or stop with its reason, every source outcome that was not clean and open questions. A missing results file is a failed step.

Team persona

The role it mirrors is presented on the team page Rosalind Wexford — Money & Influence Disclosure Analyst. For the other investigation-desk agents, see Agents and Automation: The AI That Works Here, and Its Limits and The Oversight Investigation Pipeline Agent.

Factual correction requests: If you believe information in this page is incorrect, please contact factcheck@citizenanalyst.ai with your name (optional), the specific claim, and any supporting documentation. We review all submissions and correct verified errors promptly.

Was this article helpful?
0 out of 5 stars
5 Stars 0%
4 Stars 0%
3 Stars 0%
2 Stars 0%
1 Stars 0%
5
Please Share Your Feedback
How Can We Improve This Article?