Evidence, Sourcing and Language Rules
Last updated: November 20, 2026
This is the first of the rule families in Rules, Gates and Guardrails: How We Constrain Our AI and Ourselves. It covers what we may claim, what we must cite, and what words we may not use without a court record.
For each rule we say whether it is written policy applied by editors, a warning from a program, or a block that stops a document. Most of these rules are policy. A few have a check in code.
The evidence tiers
Every factual claim in an accountability profile must trace to a source, and the source decides how the claim is worded. The four tiers are Documented, Credibly Reported, Alleged and Excluded. Our plain-language summary is in Rules & Editorial Standards — Overview, so we do not repeat the table here.
How this is applied:
- Policy. The tiers are written in
patriot-accountability-profile-standards.mdc. Editors and AI drafting tools are expected to follow them. - Informational check. When a draft contains a scored assessment,
make check-draftreports the best evidence tier behind each scored action. That check is informational only. It never fails a draft. - A separate automatic format check. Machine-generated patches from our news agent and tracker are checked for evidence format by
scripts/lib/evidence_gate.py. That check does not apply to articles written by people.
So no program reads an article and rules on whether a claim sits in the right tier. That judgment belongs to the editor.
Excluded sources, and AI is never a source
The rules exclude several kinds of material: unsourced assertions, anonymous posts, opposition-research databases without independent verification, and AI inference. Text an AI model produced is not evidence. An AI can help find a source. It cannot be the source.
One named opposition-research database is on the excluded list. The rule says never cite it and remove existing references.
In code, a mention of that database’s name produces an error-level finding that warns but does not block a publish. The reason is recorded in the code: the check matches the bare word, so it also matches documents that describe our sourcing policy. A human looks at each new hit.
Primary records first
For officials and nominees we pull the financial disclosure first. For donors we pull campaign finance filings. For corporate board members we pull the proxy statement. For nonprofits we pull the tax return. Aggregator sites are leads, never the cited source for something a primary record can carry. When a primary record cannot be fetched, the draft says it was unavailable. It does not skip it silently. This is written policy.
Wording for criminal claims
The rules say what words fit what a record shows:
- Convicted: state the crime, date, court and docket or citation.
- Indicted or charged: “was indicted on charges of,” never “committed.”
- Under investigation: say who reported it, when, and about what.
- Alleged only: say “allegedly,” name the single source, and say it is not independently corroborated.
- Association: attending the same event is not association with someone’s crime. The profile documents the specific link.
Speech is not a crime. We document what people did separately from what they said. Speech alone is never described as criminal conduct unless a court or official body made that finding.
The code backs one part of this with a block. The citation gate (scripts/lib/citation_gate.py) refuses to publish a profile that accuses a named person of a crime and carries no checkable source at all: no case number, no court-record reference and no external source link. It applies at publish. A draft can be saved uncited while the author fixes it. A per-run override flag exists for a documented exception. The gate was added after a 2026 batch of profiles shipped with accusations and no citations. The rest of the wording rules are policy that editors apply. See Publish-Time Gates.
Prohibited language
Some characterizations require a court record or official finding. The rule lists “traitor” and “treasonous,” “criminal,” “terrorist,” “coup plotter,” “pedophile” and “child abuser,” and “lied” and “lying.” For each, the rule gives a defensible alternative, such as “convicted of seditious conspiracy” in place of “traitor.”
What the code does is narrower than the rule:
- The publish-time guard audit looks for six of these words and warns: traitor, treasonous, terrorist, coup plotter, pedophile and child abuser. It does not block.
- The draft check adds “lied” and “lying,” also as a warning.
- “Criminal” is in the written rule and in the prompts given to our AI drafting tools, but no program searches for it.
The checks are heuristics, and an editor reviews every hit. The rule itself is the standard.
Unverified claims
A claim that could not be verified during drafting carries a “needs verification” tag with the date, and the rule says the tag may not be removed without a verified citation. In code, the publish audit warns about the tag and does not block. The strict mode of the draft check (make check-draft with its strict option) turns the warning into an error for a draft. Our batch builders run in strict mode.
Numbers
The numbers rule (patriot-numbers.mdc) says what every published number must carry:
- A percentage names its base.
- An “X times” comparison gives both raw figures.
- A rate states its denominator and its year.
- A dollar comparison across years is adjusted for inflation, or says it is not.
This is policy, applied with a skill called numbers-and-statistics-integrity. No code gate checks numbers.
Archive before you cite
When we cite a web page for a contested claim, we capture the page at the moment we cite it, because pages change and disappear. The rule adds two points: archive.today is never the capture of record, and a missing Wayback Machine snapshot is not proof that a page never existed.
The capture tool (scripts/ca_evidence.py capture) fetches one page, reads the site’s robots.txt first, refuses when robots.txt disallows the page unless an editor records an exception, and writes a record to our evidence store. In make check-draft, a person, organization or agency draft gets one warning when its Sources links have no capture. It is a warning only. Deciding which citations back a contested claim is still the writer’s call.
Voice rules
Voice is chosen by a document’s front matter. We keep four apart, because each one makes a different claim about who is speaking:
- Institutional voice (the Civic Accountability Format) for articles and briefs.
- Peter’s opinion voice for content flagged as Opinion.
- Substack format, an email-first layer over the right voice.
- Coaching voice for “how we build it” explainers.
These are written rules, applied by editors and by the AI drafting tools loaded with them. A shared list of AI-writing “tells” feeds the Checks panel in our editor and the drafting prompts. It warns and never blocks. See AI and Research Runtime Guardrails.
Legal sources
For constitutional law, civil rights, voting rights and the justice system, the rule says to pick at least two legal sources from different methodological or ideological starting points, rather than defaulting to the most visible commentator. This is policy.
Markdown and front matter
An article that is not in the exact Markdown subset and front-matter format our publishing tools expect can publish wrongly. Features that look fine in an editor can publish as literal text, so the rule bans them. The draft check blocks on these, with checks named obsidian-constructs and frontmatter-schema. See Draft Checks and Human Review.
What an editor still decides
An editor decides which tier a claim belongs in, whether a source is credible, which citations back a contested claim, and whether wording fits the record. The full inventory of rule files is in The Written Rules: Every Rule File and What It Governs. For how we handle who gets written about at all, see Inclusion and Fairness Gates. For the legal reasoning behind the wording rules, see Legal Framework and First Amendment Protections.
Sources
- Citizen Analyst rule files
patriot-accountability-profile-standards.mdc(May 20, 2026),patriot-numbers.mdcandarchive-before-cite.mdc(November 14, 2026), as read November 20, 2026. - Citizen Analyst code:
scripts/lib/guardrails.py,scripts/lib/citation_gate.py,scripts/lib/draft_checks.pyandscripts/lib/evidence_gate.py, as read November 20, 2026.
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.
