Draft Checks and Human Review
Rules, Gates and Guardrails

Draft Checks and Human Review

Skip to main content
Table of Contents
< All Topics
Print

Draft Checks and Human Review

Last updated: November 20, 2026

Our publish-time gates, described in Publish-Time Gates, are the last automated stop. They work better because much earlier checks catch most problems while a document is still a draft. This page covers that earlier layer, the promotion step that moves a draft toward publication, and the decisions that remain with a human editor.

For the reader-level summary of how content moves from draft to publication, see Editorial Review & Publishing Process. We do not repeat it. This page is about what the machinery does and does not enforce.

Findings while a draft is being written

Compose, our editing console, runs the same guard rules that the publish step uses every time an editor saves a document. A save still succeeds, but each finding is labelled by what it will do later:

  • BLOCKS PUBLISH for a rule that stops a publish.
  • A label naming what will refuse the document, for the two error-level rules that do not block a publish.
  • warning for the rest.

The label is derived from the same code that decides whether a publish is blocked, so what an author reads at save time and what the publish step does cannot drift apart. The point is timing. An author should learn that a document fails a gate while the document is open, not days later in a different session. The full rule list is in Publish-Time Gates.

The draft completeness check

make check-draft is the stricter, whole-document check. As of November 20, 2026 it holds 44 registered checks: 30 at ERROR severity, 12 at WARN, and 2 at INFO. An ERROR stops the draft from moving on and makes the command exit with a failure. A WARN is reported. An INFO is context only.

It first decides what kind of document it is reading (a person profile, an organization profile, an agency damage assessment, or a generic document), because the kinds carry different requirements. Some of what it checks:

  • Structure. A person profile must contain the full set of required sections, including the block that explains why the person is included and the Legal Disclaimer. An agency damage assessment must contain its own, longer list.
  • Size and sourcing. A person profile must reach 1,200 words and list at least five numbered sources. An agency damage assessment must reach 2,500 words.
  • Stubs. A document still carrying a stub tag or stub banner fails.
  • The publish gates, early. The citation check, the guard audit, the DMA and SMA designation checks, and the check that every wikilink resolves all run here as well, so a draft cannot pass this step and then fail the same rule at publish.
  • The inclusion record. The structured pre-screen block must be well formed. A recorded result other than pass is a warning, and an error under strict mode.
  • Priority tier. A hand-set tier may not exceed what the cited evidence supports. A tier below the ceiling is an editorial choice and never an error.
  • Evidence capture. A warning when web sources in the Sources section have no archived capture, in line with the archive-before-cite rule. It is never an error.

--strict promotes four warnings to errors: an inclusion result that is not pass, an outdated last_updated date, a missing priority line, and any remaining needs-verification tag. Our batch tooling runs the check in strict mode. The check never publishes anything; it reports, and a human acts on it.

Case captions must exist

make check-captions exists because of a specific failure. One batch of our October 2026 bulk run contained two fabricated court cases. Our citation gate recognizes a case caption as a citation form and stops there; it never checked that the case is real.

The check extracts every caption in a document (X v. Y and In re X) and looks each one up on CourtListener, searching opinions first and then the RECAP docket archive. Each caption ends in one of four outcomes:

Outcome Meaning
Resolved Every word of the caption appears in a result’s case name; the CourtListener link is recorded
Ambiguous Only a partial match
Unresolved Nothing close, or the lookup failed
Budget-skipped The research budget or the provider’s rate window refused the call

The command fails on any unresolved or ambiguous caption. If a caption was only budget-skipped, the run pauses with a distinct exit code, and a skipped lookup is never treated as a pass. CourtListener limits each account to 10 requests per minute, 75 per hour and 300 per day, so the lookups share one limiter and one cache. An optional flag adds the priority-tier check to the same run.

Promotion: draft to review

A draft lives in a staging folder that no publish channel reads. After a human reviews it, make promote SLUG= moves it toward publication. Promotion:

  1. Validates required frontmatter and runs the guard audit. Any error-level finding aborts the promotion and the draft stays where it is. That is stricter than publish, which exempts two error rules from blocking.
  2. Refuses a second published document that claims the same subject and facet.
  3. Files the document under its category and sets its status to review.
  4. Repairs reciprocal related: links and updates the index.

review still does not publish. An editor sets status: published, and only then can a publish run push the document live. The promotion tool does have an option that sets the status to published in one step, but it is an explicit choice by the person running it. Our written rule is that no AI session runs the publish step on its own.

What a person always decides

Code can check whether a block exists, a number is present or a case is real. It cannot decide what is right. These steps stay with the editor, and no check substitutes for them:

Decision What the machine does What the human does
Who is included Validates the pre-screen record, refuses a non-pass result at publish Decides inclusion or exclusion, and applies the private-citizen test
Criminal-claim wording Blocks a charge with no checkable source Approves the exact wording, and whether a claim is a finding or an allegation
Priority tier (P0, P1, P2) Computes a ceiling and rejects a tier above it Chooses the tier
Final publish Gates, then waits Sets published and runs the publish
Every correction Detects a changed factual claim and files it as pending Reviews it, then confirms or dismisses it

Our research and drafting agents are bound by the same line. Their definitions say they never commit, push or publish. They leave a draft or a findings file for a person to read. The policy is in How Citizen Analyst Uses AI; the correction step is in Post-Publish Verification and Corrections.

What these checks do not do

They do not make a claim true. A draft can have every section, five sources and a real case citation, and still be wrong. That is why the human steps exist.

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?