AI and Research Runtime Guardrails
Rules, Gates and Guardrails

AI and Research Runtime Guardrails

Skip to main content
Table of Contents
< All Topics
Print

AI and Research Runtime Guardrails

Last updated: November 20, 2026

How Citizen Analyst Uses AI says what we commit to about AI. This page says how those commitments are held while the tools are running. Some are enforced in code, and some rest on written rules applied by people. For each one we say which.

AI output is never evidence

This is the rule everything else hangs on. Our evidence standard places “AI inference” in the Excluded tier: it must not appear in a profile. Search results and drafted text from AI tools are leads for an editor to check against a court record, an official document or credible reporting.

This is a written standard applied by editors. No code can tell whether a sentence started life in a model. What the checks in Draft Checks and Human Review and Publish-Time Gates can do is demand that claims carry sources a person can open.

One gateway for research

Research tooling goes through one command-line tool, scripts/ca_research.py. Our research agents are defined to use it and nothing else: their definitions say they never make a network request any other way. That is a rule in each agent’s written instructions, not a firewall.

The gateway gives every research call the same properties:

  • Cached. A repeated request is answered from a local cache and costs nothing.
  • Budgeted. A run has a hard budget for each billable kind of call: web search, court-record lookups, page fetches and official-API calls. Before a billable action runs, the tool charges the budget. If the kind has nothing left, the action never runs and the tool exits with a dedicated code. A budget for a running job can be lowered but not raised without a written reason in the run’s ledger.
  • Metered. When a provider’s rate window is full, the tool says how long to wait and does not retry on its own.
  • Recorded. Each call writes a row to a ledger: the tool, its arguments with secrets removed, the outcome, the cost and whether it was a cache hit.
  • Classified. A failure is labelled (a timeout, a quota, a parse error), so an agent records an unverified lead rather than guessing past the gap.

Every line the tool prints is scrubbed of secret values. An agent that runs out of budget or hits a rate limit records that as a finding and stops that line of research.

Metered APIs

Some services we use are metered by their owners. CourtListener limits each account to 10 requests per minute, 75 per hour and 300 per day. The limit is per account, not per program, so a lookup from our editing console, a command-line session and a scheduled script all draw on one allowance. Our client keeps one shared limiter and a disk cache for these calls, and the case-caption check uses the same limiter and the same research budget. The limiter coordinates callers on one machine only, so the cap CourtListener enforces remains the final limit.

Archive before cite

A web page cited for a contested claim can be changed or removed by its publisher after our story runs. Our rule is that such a source is captured when we cite it.

  • The capture command, scripts/ca_evidence.py capture, fetches that one page and follows no links. It reads the site’s robots.txt first and refuses a disallowed page, unless an editor records a public-interest exception.
  • It writes a record to our committed evidence store: the HTTP status, final address, a content hash, the title and a short excerpt. It records the nearest existing Internet Archive snapshot, and submits a new one only when that is configured. Otherwise it prints the manual step.
  • It refuses archive.today in every form, because that site is not a capture of record we trust.
  • A missing archive snapshot is never evidence that a page did not exist.

The enforcement is soft. make check-draft gives a warning, never an error, when a profile’s Sources section holds web addresses with no capture. Deciding which citations back a contested claim is still the writer’s call.

AI images carry a label

Every image our editing console generates with an AI model gets two disclosures:

  1. A visible “AI-generated illustration” band burned into the pixels. It survives a screenshot, a resize or a re-post, because it is not metadata.
  2. A C2PA Content Credential, a signed record that a model made the image.

Both are applied at one choke point that every generated image passes through. The credential is self-signed, so a verifier reports the signer as unrecognized. Any crop, resize or re-save drops the credential, which is why the label is in the pixels. If signing is not configured, the image still receives the label, goes out unsigned, and a warning is logged.

A third rule has no code behind it: no AI image of a real, identifiable person in an invented scene. A picture of a real person doing something they did not do is a false claim about them, and a label does not make it true. No software reads an image for faces. The prompts our tools use for covers and slideshows ask the model for no faces and no real people, and an editor checks before an AI image goes into a document.

Confidential audio stays on the machine

A source’s recording is the source. Source audio is transcribed locally only, with scripts/ca_transcribe.py, which has no cloud backend. A test fails the build if either file imports an HTTP or cloud library. Transcription runs offline, and a source transcript is written only to a path that version control ignores. Public-record audio, such as hearings and press conferences, may use a cloud service.

The tool enforces this for itself. Nothing stops a person from uploading a file to a website; choosing the safe path is the reporter’s job.

Scraping conduct

When a tool collects data from a website, our written rule is to honour robots.txt and the site’s terms unless an editor records a public-interest exception before collection runs. Tools throttle their requests, stop at a login page, and keep only the fields a story needs. No code checks robots.txt or terms in general; the page-capture command above is the exception, because it reads robots.txt and refuses. The method and the legal background are in Scraping Ethics and Methods.

The public Rights Advisor

The Rights Advisor is the chat assistant on our site. Its limits come from its server code:

  • The key stays on our server. The request to the AI provider is made by our server, never your browser.
  • Fixed modes. A request must name a mode from a short fixed list, and any persona must come from another. It must contain a message, and the last message must come from the user. Anything else is rejected.
  • Grounded in our articles. The user’s question is used to retrieve relevant articles from our knowledge base, and excerpts go to the model with it. The reply includes a list of the articles retrieved for the question.
  • System-prompt limits. The instructions tell the model to cite specific laws and cases, to recommend consulting an attorney, to favour nonviolent legal approaches, never to give advice that could facilitate illegal activity, never to assume a user’s identity or citizenship status, and never to store or share personal information. These are instructions to a model, not guarantees enforced by code.
  • Refusals are surfaced. If the model declines, the server returns an error, not a half-answer.

Zero-PII logging. The request log records only the method, the path, the status and the time taken. It does not record an IP address, a user agent or a request body. A second layer redacts anything shaped like an IPv4 address, an IPv6 address or an email from every value written to a log. The chat code does not save conversations. No account is needed. As our AI-use policy states, the Advisor gives general information, not legal advice.

For how the Advisor fits with our other agents, see The Rights Advisor AI.

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?