The Profile Builder Agents and the Batch Kit
Agents

The Profile Builder Agents and the Batch Kit

Skip to main content
Table of Contents
< All Topics
Print

The Profile Builder Agents and the Batch Kit

Last updated: November 20, 2026

Accountability profiles are our most carefully governed documents. Two agents help draft them, and a tool called the batch kit lets us run many at once. All produce drafts; none publishes. Bulk scale changes how many we dispatch, never how well each is sourced.

What it is

Two builder agents and one tool.

  • compose-profile-pipeline builds or advances one accountability profile from start to finish, reproducing in explicit stages what our editing console does. It mirrors the team’s profile builder: someone who researches, drafts to the template and walks the draft through every gate.
  • bulk-profile-builder finishes one profile draft, from a stub to a complete document, inside a batch wave. It is dispatched by the batch kit and by nothing else.
  • The batch kit (scripts/ca_batch.py) organizes a wave: it seeds the rows, stages checks, hands out batches, records each agent’s outcome and commits finished drafts.

Both follow the published skill Accountability Profile Builder.

When it runs

compose-profile-pipeline runs on demand when a human asks to build, draft, gate or check a profile for a named person or organization. Before it scaffolds anything it asks for the subject, slug, category and classification (official, private citizen, spokesperson or federal judge). It does not guess these.

bulk-profile-builder runs only when the batch kit dispatches it. In the kit’s code, one dispatch covers one to three profiles from the same cluster, each its own complete document, and at most eight dispatches are in flight at once. The prompt is frozen when a run starts; the kit refuses to render a batch from a prompt that changed mid-run.

What it does

The single-profile pipeline

  1. Scaffold. It writes a stub in _drafts/ with status: draft, which stays draft throughout. An existing profile is edited in place, never copied into a dated sibling.
  2. Pre-screen. Before any research it runs the pre-screen (scripts/ca_prescreen.py). A subject that fails, such as an excluded subject or a private citizen without a qualifying anchor, stops the work.
  3. Research. A deterministic pre-fetch gathers the records the subject’s class requires into a committed evidence store. The agent then researches only through scripts/ca_research.py under a budget. Primary disclosures come first: financial disclosures for officials and nominees, campaign-finance filings for donors, proxy statements for board members, Form 990 for nonprofits. OpenSecrets and Wikipedia are leads, never the cited source.
  4. Draft. It follows the profile skill’s section order exactly, because later gates look for blocks in fixed places. Every factual claim is attributed to a source.
  5. Eligibility gates. A private citizen must pass the inclusion gate. A federal judge qualifies only for personal misconduct, never rulings alone. A spokesperson faces a tenure-window rule before any scoring.
  6. Scoring gates. It applies the scoring instruments that fit the subject, following Scoring Instrument Rules: Citizen Analyst — Malice Evaluator (Democratic Malice Assessment) for actions, Citizen Analyst — Spokesperson Misinformation Assessment (SMA) for a spokesperson’s statements, Citizen Analyst — Dangerous Rhetoric Assessment (DRA) for a public figure’s attributable speech. The Democratic Malice Assessment is scored only for nominees who have been confirmed and have served. An unconfirmed, pending, returned or withdrawn nominee carries a block that reads “not scored: no actions in office.”
  7. Credibility, citation and wording checks. Source Credibility Scoring (Document Credibility Score), Citation Checker and Citizen Analyst Sanity Check are applied, with attribution and derivative-work checks done by reading. Assessment sections end with the marker that renders our legal disclaimer at publish, and loaded labels that need a court record are checked.
  8. Validation. It runs the dry-run promotion check, the strict draft gate (scripts/check_draft.py --strict) and the case-caption check (scripts/check_captions.py). A caption it cannot resolve is fixed or removed.
  9. Stop. Promotion happens only if the human asks. This agent never publishes.

It works in its own isolated copy of the repository and commits each checkpoint by explicit file path on its own work branch. It never pushes, merges, opens a pull request or publishes.

The bulk builder and the batch kit

The bulk builder reads the prompt the kit rendered for its batch and finishes each named profile to the same standard. The strict draft gate must pass for every row first. Its results file gives, per profile, the priority, the scoring result or “not scored”, the source count, its own credibility score, its sanity-check verdict, the gates it applied by judgment and any items left tagged for verification.

The kit itself works this way:

  • State is derived, never hand-set. Each call recomputes every row’s state from the seed list, an event log and the draft files themselves, so a session that dies mid-wave loses nothing.
  • Failures are explicit. A row failing a check is recorded with a reason. A crash stops the run and a refused research budget pauses it; neither hands out more work until a human resumes it.
  • Closing a batch commits drafts by explicit path. The kit refuses to commit in the main checkout or on the main branch. Pushing and merging are not in the kit at all.

Skills it applies

Published skills: Accountability Profile Builder, Citizen Analyst — Private Citizen Inclusion Gate, the three scoring methods above, Source Credibility Scoring (Document Credibility Score), Citation Checker and Citizen Analyst Sanity Check.

Internal working skills, described in words only: a check on how much quoted text is acceptable, a check that each claim traces to where it came from, and the batch-run procedure for a wave.

Rules and gates it obeys

Most of the pipeline is editorial discipline written as skills, not a checker that blocks. The inclusion gate, scoring instruments, sanity check and citation chain are procedures the agent must actually carry out. Its final report must say which gates it applied by judgment rather than a code check, so a reader knows which parts are an editorial call and which are a verified pass.

What it may never do

  • Push, merge, open a pull request, run the GitHub tool or publish. The bulk builder also never stages or commits, and never creates or switches a branch.
  • Promote a draft or change its status. The single-profile pipeline promotes only when asked, and never sets status: published.
  • Write to WordPress, the search index or a server.
  • Touch any file other than the rows it was given: no other row’s draft, manifest, skill, rule file or project instructions.
  • Fabricate a source, quote, date, docket number or statute. A claim it cannot source is dropped or tagged for verification.
  • Make a network request except through scripts/ca_research.py, or read the file that holds our credentials.
  • Mark a gate passed without doing the skill’s work.

Where a human decides

A human gives the single-profile agent its inputs and decides whether to promote and publish. In a wave, the kit commits drafts only after its own check stage passes, and an orchestrator spot-reads one document per batch. An editor promotes each profile and sets it to published. A profile that cannot reach standard is reported as partial or blocked, never shipped thinner.

Team persona

The role these agents mirror is presented on the team page Mara Castellano — Senior Profile Architect. Independent checking of finished drafts is described in The Research Worker and the Document Auditor, and the full set of agents is in Agents and Automation: The AI That Works Here, and Its Limits.

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?