Rules, Gates and Guardrails: How We Constrain Our AI and Ourselves
Rules, Gates and Guardrails

Rules, Gates and Guardrails: How We Constrain Our AI and Ourselves

Skip to main content
Table of Contents
< All Topics
Print

Rules, Gates and Guardrails: How We Constrain Our AI and Ourselves

Last updated: November 20, 2026

Citizen Analyst uses AI to research, draft and check work, and it publishes about real people. That combination needs limits, and the limits need to be visible. This page is the entry point to a set of articles that describe those limits as they exist in our code and our written rules today.

We try to be exact about enforcement. Where a program stops something, we say it blocks. Where it only warns, we say it warns. Where a rule is written policy that an editor applies by judgment, we say that too. Our AI tools and our team pages are bound by the same rules. The specialists listed in Citizen Analyst — Organization Chart are AI personas, and the agents described in Agents and Automation: The AI That Works Here, and Its Limits run inside these limits.

Three words, three meanings

  • A rule is a written standard. It says what we will and will not do, such as “never cite a people-search result.”
  • A gate is a check that can stop work. A publish gate refuses to publish a document. A draft gate refuses to let a draft move forward.
  • A guardrail is a warning or a limit that shapes work without stopping it. A flag in the editor, a warning in a publish report and a spending budget are all guardrails.

Many rules have no gate behind them. We do not hide that. It is why a human review step sits at the points that matter most.

The life of an article

  1. Authoring. Our AI tools load the written rule files that match the files being edited, and a skill router points the writer to the right method. Our research agents run research through one command-line tool, with a hard per-run budget and a committed evidence record. See Evidence, Sourcing and Language Rules and AI and Research Runtime Guardrails.
  2. Draft checks. The editor flags findings as a document is saved. make check-draft runs 44 registered checks as of November 20, 2026, and a separate check confirms that every court case caption exists on CourtListener. A failing check stops that stage until it is fixed. See Draft Checks and Human Review.
  3. Human review. An editor reads the draft and decides about inclusion, wording of criminal claims and priority. make promote validates the draft and runs the guard checks before it moves, and it sets the status to review. That status does not publish anything.
  4. Publish gates. Only a human starts a publish. The scripts then apply gates before anything is written to the site. There are four outbound channels: knowledge-base articles, the Substack export, generated tracker pages and timeline stories. Each gate must exist on every channel or carry a recorded reason for its absence, and a test fails otherwise. See Publish-Time Gates.
  5. Live read-back. After a publish, make verify-live reads the page back from the live site and compares it with what we sent. It reads only, and it never blocks. See Post-Publish Verification and Corrections.
  6. Corrections. A substantive correction is recorded in a ledger. A human confirms each one, and the notice is rendered onto the page from the ledger. See Corrections & Transparency Policy.

The families

Evidence, sourcing and language rules. What we may claim and how we word it: the evidence tiers, criminal-claim wording, the prohibited-language list, the numbers rule and archive-before-cite. Most are policy, with a few checks in code. See Evidence, Sourcing and Language Rules.

Inclusion and fairness gates. Who we write about at all: the private-citizen test, the federal-judge threshold, the Basis for Inclusion block, the people-search ban and the retribution tracker’s naming rule. The pre-screen result is enforced at publish. See Inclusion and Fairness Gates.

Scoring-instrument rules. The six assessment instruments, what each one scores, the check that a designation must follow from its own scores, and the mandatory legal disclaimer. See Scoring Instrument Rules.

Publish-time gates. The citation gate, the inclusion gate, the guard audit, the related: link check and the warning-only reuse-terms notice, along with the bypass flags and what they are allowed to skip. See Publish-Time Gates.

Draft checks and human review. The earlier layer, plus the steps an AI cannot skip. See Draft Checks and Human Review.

Post-publish verification and corrections. What happens after the write, including the corrections ledger. See Post-Publish Verification and Corrections.

AI and research runtime guardrails. How our AI and research tools are constrained: AI output is never evidence, spending budgets, image labels, local-only transcription of confidential audio, and the public Rights Advisor’s limits. See AI and Research Runtime Guardrails.

Engineering and change-control guards. The guards that stop the platform from breaking the rest: protected files, commit-time checks, required tests and the one-editing-device rule. See Engineering and Change-Control Guards.

The catalog of rule files. Every one of the 33 rule files, what it governs and what content it applies to. See The Written Rules: Every Rule File and What It Governs.

Why gates differ from warnings

A gate that blocks too often gets bypassed, so we are deliberate about which findings block. In the publish audit, an error-level finding blocks a publish unless the code records a reason it should not. Two error-level findings carry such a reason and publish with a warning. Bypass flags exist for documented exceptions. The guard audit’s override still reports what it let through. A bypass given for one document is cleared for any other document the same run has to republish.

We also learned that guards fail quietly, so we test them. A guard needs a test that proves it fails on bad input. That is covered in Engineering and Change-Control Guards.

What a human always decides

  • Who is included in a profile, and who is not.
  • Whether wording about a crime fits the record.
  • A profile’s priority tier.
  • Whether to publish. No AI agent publishes on its own, and publishing always starts with a human.
  • Whether a correction is warranted. Detection is automatic. The decision is human.
  • Any change to a written rule. An AI cannot change these files without a human instruction naming the change.

This matches the commitments in How Citizen Analyst Uses AI and the process in Editorial Review & Publishing Process. Our plain-language editorial standards are in Rules & Editorial Standards — Overview.

A note on honesty

Some of these rules are stronger on paper than in code. We say so where we know it, and we would rather correct our description than overstate the protection. If you find a place where we describe a rule more strongly than it is enforced, tell us at the address on our corrections page.

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?