Scoring Instrument Rules
Rules, Gates and Guardrails

Scoring Instrument Rules

Skip to main content
Table of Contents
< All Topics
Print

Scoring Instrument Rules

Last updated: November 20, 2026

We publish six assessment instruments. Each one scores a different kind of thing, and each comes with its own written rule, a methodology, and a legal disclaimer. This page is the rules view, the third family in Rules, Gates and Guardrails: How We Constrain Our AI and Ourselves. For the full toolset, see Scoring & Inclusion Systems — Complete Inventory.

Why the six are not interchangeable

A score only means something if you know what was scored. Putting a person’s actions, a person’s statements, a market and an agency on one scale would produce numbers that look comparable and are not. So each instrument has its own unit of analysis.

The central list lives in one registry file, scripts/lib/instruments.yaml. It records, for every instrument, what it scores, its section heading, its governing rule, its methodology file and whether the legal disclaimer is required. Adding a seventh instrument means adding an entry there, and a test fails until you do.

Instrument What one score is of Governing rule file Methodology (in our repository)
DMA, Democratic Malice Assessment A person’s actions patriot-malice-evaluator.mdc docs/democratic-malice-methodology.md
SMA, Spokesperson Misinformation Assessment A spokesperson’s statement veracity, limited to their time in the role patriot-spokesperson-misinformation-assessment.mdc docs/spokesperson-misinformation-methodology.md
DRA, Dangerous Rhetoric Assessment A public figure’s attributable public speech patriot-dangerous-rhetoric-assessment.mdc docs/dangerous-rhetoric-methodology.md
LIA, Legislative Impact Assessment One action (a bill, executive order or rule), on two tracks, never one number patriot-legislative-impact-scoring.mdc docs/legislative-impact-methodology.md
WLM, Policy Winner-Loser Map A market across many actions. It consumes LIA output and never re-scores it patriot-policy-winners-losers.mdc docs/policy-winners-losers-methodology.md
ADA, Agency Damage Assessment One federal institution. It consumes DMA and LIA output and never re-scores it patriot-agency-damage-assessment.mdc docs/agency-damage-assessment-methodology.md

Each instrument has a published skill that tells our AI tools and editors how to run it: Citizen Analyst — Malice Evaluator (Democratic Malice Assessment), Citizen Analyst — Spokesperson Misinformation Assessment (SMA), Citizen Analyst — Dangerous Rhetoric Assessment (DRA), Citizen Analyst — Legislative Impact Scoring (LIA), Citizen Analyst — Policy Winners & Losers Mapping (WLM) and Agency Damage Assessment. The public methodology articles are The Democratic Malice Methodology and The Spokesperson Misinformation Methodology.

What each rule requires

The rule files are written policy. They say when an instrument must be run and what safeguards apply.

DMA. Required for every Trump-administration profile (either a designation or an explicit “no designation” block), for any profile rated P0 or P1, and when an update adds documented actions. The ideology-versus-malice gate runs first: legitimate policy through legitimate process is always ideology, never malice. Higher scores need stronger evidence. Intent must be documented separately, and inferring it from outcomes caps a score at 1. Speech alone never qualifies for a DMA score. The rule adds that private citizens cannot receive the top score without a criminal conviction.

SMA. It applies to spokesperson profiles. A tenure-window guard comes first: only statements made while the person held the role count, and commentary before or after the role is excluded. Every scored statement needs an institutional fact-check citation. A “False” rating alone caps a statement at 1, because the person’s knowledge must be documented separately for higher scores. Vigorous advocacy is never misinformation. The comparison baseline is a band, never a computed “lie rate.”

DRA. It applies to public figures only. Private citizens are excluded. It scores rhetoric on three separate axes: dehumanization, targeting and incitement risk. A verbatim quote is required, and a paraphrase is never scored. The rule forbids asserting that speech is legally “incitement” or that it caused violence.

LIA. It runs a fiscal and distributional track and a rights track side by side, and never combines them into one composite score, meter or grade. The rights track reports exposure and litigation status, never a verdict that an action is unconstitutional. Estimates from other organizations are attributed to them, with the assumptions behind each.

WLM. It has two gates that both fail closed. First, benefit is not corruption: a firm that gains from a policy has done nothing wrong on that fact alone. Second, no investment advice, ever. It also bars composite scores and firm rankings, and every row must say what the market looked like before the action.

ADA. It scores an institution, not a person. Recoverability is a second axis and is never folded into the composite. The composite is withheld below 70% dimension coverage, and a stub carries no scores at all. Personnel ledgers name only senior officials or people already publicly reported. Everyone else appears in aggregate counts.

Publishing is always a human step. The WLM and ADA rules also require a claims-integrity review before anything goes live, and the LIA rule requires one for any “likely unconstitutional” call or any estimate we produce ourselves.

The check that a label must match its own scores

A designation is a label attached to a named person, so it has to follow from the scores printed next to it.

  • DMA. The ladder is fixed in code (scripts/lib/dma_ladder.py). “Pattern of Democratic Malice” requires three or more actions at score 2 or higher, or one at 4 or higher. “Sustained Campaign of Democratic Destruction” requires five or more at 3 or higher, or two at 5. Below that, there is no designation.
  • SMA. The same structure applies with different labels (scripts/lib/sma_ladder.py): “Pattern of Strategic Misinformation” and “Sustained Disinformation Campaign.”

In scripts/lib/guardrails.py, a stated DMA or SMA designation that its own scores do not compute to is an error that blocks publication, and so is a scoring section whose scores cannot be read. A missing designation line only warns. One SMA case is softer. The skill’s “Sustained” label has a trigger no score list can show, which is a documented pattern of repeating a claim after a correction. So a “Sustained” label that scores support only as “Pattern” is a warning for a human to confirm, not a block.

A published designation that has to change is a correction and goes through our corrections process, not a silent edit. See Post-Publish Verification and Corrections. For DMA, the list of documents allowed to disagree with the ladder is empty. A test keeps it empty, and adding an entry is a reviewed decision.

LIA, WLM and ADA have no label ladder to compute. Their rules govern structure instead. The draft check enforces some of it for ADA: required sections, a 2,500-word floor and the rule that a stub’s score fields are all empty. These block a draft.

The mandatory disclaimer

A document that contains one of the six instrument sections must carry a Legal Disclaimer. The missing-defamation-guard rule blocks publication without it. The trigger is the assessment heading, not a score, because a reader sees the assessment whether or not anything was scored. The disclaimer is not typed into articles. Authors insert a marker, and the registered text is rendered at publish from one source. The check accepts the marker, the disclaimer heading, or the names of the court cases the text cites. A surname that merely appears in a firm’s name does not count. The legal reasoning is in Legal Framework and First Amendment Protections.

Stored scores and the credibility score

The draft check also blocks a draft whose stored score record disagrees with a fresh read of its own prose. That keeps a number in the data from drifting from the number in the text.

A separate Document Credibility Score measures sourcing rigor, not political direction. It is an editor’s tool. It is not one of the six, it advises by default, and it blocks a publish only when someone sets a minimum for that run.

What a human decides

People decide whether an instrument applies, whether an action is ideology or malice, whether a statement is advocacy or misinformation, and whether to publish. An AI tool applies a rule file. It does not own the judgment. The catalog is in The Written Rules: Every Rule File and What It Governs, and the inclusion tests that come before scoring are in Inclusion and Fairness Gates.

Sources

  1. New York Times Co. v. Sullivan, 376 U.S. 254 (1964), Milkovich v. Lorain Journal Co., 497 U.S. 1 (1990) and Brandenburg v. Ohio, 395 U.S. 444 (1969), the cases cited in the disclaimer text.
  2. Citizen Analyst rule files for the six instruments and scripts/lib/instruments.yaml, as read November 20, 2026.
  3. Citizen Analyst code: scripts/lib/guardrails.py, scripts/lib/dma_ladder.py, scripts/lib/sma_ladder.py and scripts/lib/draft_checks.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.

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?