Scoring Framework

Methodology

The complete methodology: what we measure, how we score it, and what we don't claim to know.

Rubric v1.0 · 14 dimensions · Fixed, versioned model configuration

What Undersignal measures: The rhetorical mechanics of written content — the structural techniques that shape how information is processed, independent of whether the information itself is true or false. We score how content is constructed to influence belief formation, not whether you should agree with it.

What We Measure

Undersignal analyzes written content for rhetorical mechanics — the structural techniques that shape how information is processed, independent of whether the information itself is true or false.

We do not measure:

We measure how content is constructed to influence belief formation. A factually accurate article can score high if it uses fear amplification, manufactured consensus, or identity pressure to drive conclusions. A factually dubious article can score low if it simply states claims without rhetorical engineering.

The score reflects structural architecture, not truth value.

What You Can Submit

URLs. Any publicly accessible webpage containing primarily text content. News articles, blog posts, opinion pieces, press releases, research summaries, newsletters.

Direct text. Paste raw text for analysis when URL fetch isn't possible or you want to analyze specific passages.

PDFs — upload directly via the PDF icon in the analysis interface. Text-based PDFs only; scanned/image PDFs without extractable text cannot be analyzed.

Speeches & transcripts — paste the transcript text directly or submit a transcript URL.

What we cannot score

Video content (no transcript extraction)
Audio content (podcasts, radio)
Scanned or image-only PDFs without extractable text
Paywalled content we cannot fetch
Social media posts shorter than ~200 words (insufficient structure to analyze)

How Scoring Works

The 14 dimensions

Every piece of content is evaluated across 14 dimensions: 13 scored on a 1 to 10 scale, plus the Disguise Factor on a 0 to 3 scale. Each dimension has defined anchor points — specific observable characteristics that map to score values.

Fear Amplification
2 No existential or mortality framing; stakes proportionate to evidence
5 Fear present but grounded in evidence; threat real but not inflated
9 Existential or mortality stakes invoked repeatedly and disproportionately; survival anxiety activated to bypass analytical processing
10 TMT-level mortality salience; piece frames the issue as civilization-ending with no evidential basis
Authority Exploitation
2 Experts cited accurately within their expertise; uncertainty acknowledged
5 Experts cited selectively; opposing views excluded
9 Authority weaponized to assign certainty to uncertain questions; experts' conclusions applied beyond their scope
10 Authorities explicitly deployed to foreclose inquiry on questions they did not study
Tribal Signaling
2 No identity signaling; reader's identity irrelevant
5 Tribal architecture present but implicit
9 Holding alternative view carries explicit social, professional, or moral cost
10 Tribal enforcement is the thesis; dissent explicitly equated with moral failure or complicity
Emotional Loading
2 Neutral, calibrated tone
5 Mild emotional loading; word choices carry valence
9 Emotion deployed as substitute for evidence, disproportionate to facts
10 The piece contains no argument that survives emotional removal
Manufactured Consensus
2 No consensus claims; piece acknowledges genuine debate
5 Some social proof deployed; "many experts say" without naming dissenters
9 False unanimity asserted; alternative views excluded or dismissed
10 Consensus manufactured through circular citation or documented coordinated messaging
Centralized Narrative
2 Facts support multiple interpretations
5 Preferred interpretation present but alternatives acknowledged
9 Architecturally closed; conclusion pre-stated in headline; body confirms verdict
10 Every structural element works in concert to foreclose all alternatives
Identity Pressure
2 No cost to disagreement
5 Soft, implicit pressure
9 Disagreement carries explicit reputational or moral cost
10 Disagreement carries named, specific consequences; cost of dissent is the piece's central argument
Archetypal Framing
2 No character roles; institutions are actors
5 Implicit characterization
9 Archetype assignment is the primary persuasive mechanism
10 Character assignment entirely replaces argument; hero/villain structure is the piece's only logical mechanism
Money Context
2 Beneficiaries obvious, disclosed, or irrelevant
5 Beneficiaries exist but unnamed
9 Narrative directly serves financial interests of actors absent from or cast as neutral
10 Undisclosed financial relationship between publisher/author and story subject is directly traceable
Logical Fallacies
2 Claims sourced, qualified, accurate
5 Minor framing biases; core claims defensible
9 Entire argument rests on conflation or misapplication of evidence
10 The central logical claim is internally self-defeating when stated explicitly
Rapid Compliance
2 Low adoption potential; requires prior knowledge
5 Moderate; framing adoptable but requires engagement
9 Highly compressible; entire argument fits in headline, meme, or tweet; designed for rapid spread without reading
10 Explicitly engineered for memetic replication; replication requires zero comprehension
Suppression & Promotion
2 Reports what happened; no alternatives to suppress
5 Selective emphasis through story placement
9 Active disqualification — competing hypotheses labeled conspiracy, fringe, or debunked
10 Suppression with documented real-world consequence
Narrative Structure
1 Pure list of claims; no story form, no characters
4 Partial story elements but structure is loose
7 Clear story arc (setup → conflict → resolution); reader is transported
10 All narrative roles filled; complete emotional arc; reader has no analytical foothold outside the story
Disguise Factor 0–3 scale
0 Fully transparent
1 Mild ambiguity (byline missing, native ad disclosure buried)
2 Active misrepresentation (fake news site designed to look like mainstream outlet; PR as journalism)
3 Deliberate identity fraud (impersonation of a real outlet, fabricated author, state-sponsored content as organic journalism)

Clusters and weights

Dimensions are grouped into six clusters. Within each cluster, only the highest-scoring dimension contributes — we take the maximum, not the average.

Identity Exploitation Identity Pressure, Tribal Signaling
Emotional Activation Fear Amplification, Emotional Loading, Archetypal Framing
Legitimacy Manipulation Authority Exploitation, Manufactured Consensus, Money Context
Narrative Control Suppression & Promotion, Centralized Narrative, Narrative Structure
Behavioral Pressure Rapid Compliance
Logical Integrity Logical Fallacies

Clusters are weighted by their documented impact on audience persuasion. Identity and emotional mechanics carry the highest weight; logical integrity carries the lowest.

Score calculation

Dimension scores feed into a proprietary weighted formula that emphasizes the intensity of deployed mechanics over breadth. A single dimension operating at extreme intensity can produce a high score. Content that deploys mechanics across many dimensions receives an additional breadth signal, but intensity is the primary driver.

The formula works in three stages:

  1. Cluster scoring. Within each cluster, only the highest-scoring dimension contributes. This prevents low-level activation across many dimensions from substituting for genuine intensity.
  2. Weighted aggregation. Cluster scores are combined using proprietary weights that reflect each cluster's documented impact on audience persuasion and cognitive processing. Clusters with greater manipulation potential carry higher weight.
  3. Disguise amplification. The Disguise Factor acts as a multiplier on the final score. Covert persuasion — content that conceals its persuasive intent — scores materially higher than overt persuasion at the same mechanic intensity. A disguise factor of 2 or 3 will substantially elevate a score that would otherwise be moderate.

Scores are normalized to a fixed 1.0–10.0 scale calibrated against real content. The full range is genuinely used: 1.0 represents the floor of structurally non-rhetorical reference content; 10.0 represents the ceiling of content engineered across multiple dimensions with maximum disguise.

Boundary rule

When evidence supports both N and N+1 equally, assign N. This conservative default prevents score inflation and ensures consistency.

Nonpartisan symmetry

We score rhetorical function, not vocabulary or political direction. The analysis verifies: "Would I score identically if this structure appeared in content from the opposite political orientation?" If no, the score is adjusted.

Construction Labels

Construction labels describe the structural architecture of content based on its dimension score patterns. They are derived from scores after analysis — not independently assigned by the model, and not used to constrain or adjust the score. A construction label tells you how the content is built. The score tells you how intensely.

A blog post can score 4.2 and be labeled Structured because it was deliberately crafted. Construction and score are independent.

Declarative

Structurally non-rhetorical; states facts without arguing a position.

Neutral

Some mechanics present but not concentrated; mild rhetorical presence.

Patterned

Organic patterning, not deliberately constructed.

Elevated

Deliberate mechanics with visible intent.

Structured

Deliberately constructed framing with low transparency. The Disguise Factor is what separates Structured from Elevated.

What Scores Mean in Context

Context matters. A 5.0 in a political op-ed is unremarkable — advocacy framing is that format's job and the intent is declared. A 5.0 in a corporate earnings call warrants scrutiny. A 5.0 in an academic paper is unusual and worth examining. The construction label and intent verdict provide the context the raw score alone doesn't carry.

Scores by content type, for orientation:

1.0 – 3.4 MINIMAL Structurally non-rhetorical or minimally shaped content. Legal filings, academic abstracts, government statistics, wire news briefs.
3.5 – 5.4 MODERATE Moderate rhetorical mechanics present. Standard editorial framing, advocacy journalism, declared opinion, corporate communications.
5.5 – 7.4 HIGH Heavy editorial framing, strong persuasion mechanics, partisan content, commercial influence.
7.5 – 10.0 CRITICAL High-intensity mechanics across multiple dimensions. Often includes elevated disguise factor. State media, coordinated influence operations, engineered disinformation.

The product displays Minimal, Moderate, High, and Critical as severity labels on every report. The scale marker shows: 1 · Minimal, 5 · Moderate, 10 · Critical.

What We Don't Claim

We do not determine whether claims are true or false.
We do not assess legal intent or culpability.
We do not make judgments about the author's character or motivations (intent is assessed separately in the intent verdict).
We do not determine whether content caused harm.
A high score is not an accusation. It is a description of structural mechanics.
We do not score entities. A score describes one submitted text; naming its outlet or author is attribution for that text, not a verdict on the entity.

Frequently Asked Questions

Why does the same URL sometimes produce different scores? +

Two situations:

1. If you've submitted this URL before, you're seeing the saved result. Reports are saved and versioned, keyed on the content itself, so resubmitting the same content returns the report on file rather than a fresh run.

2. First-time submission of a dynamic page (major news sites, live pages): our fetch pipeline captures the page at a moment in time. Dynamic pages can load different related articles, sidebars, and ads depending on timing, so the captured text can differ between fetches. Pasting the article text directly gives you control over exactly what is scored.

What happens when the URL fetch fails or returns noisy content? +

We use four fetch methods in priority order: Firecrawl (JS rendering, primary) → ScrapingBee (residential proxies) → Jina.ai Reader → direct HTTP.

If all methods fail, we return a content-blocked message and ask you to paste the text directly.

If a lower-priority fallback succeeds but returns noisy content — navigation links, ad copy, related article headlines alongside the actual article — dimension scores may be affected. The system will score what it receives. For best results on paywalled, heavily dynamic, or ad-dense sites: paste the article text directly instead of submitting the URL.

Does the outlet identity affect the score? +

No. The outlet is identified and reported in the source context section, but outlet identity does not affect dimension scores. A Fox News article and an MSNBC article using identical rhetorical mechanics are scored by the same rules with no source adjustment. We score what the text does, not who published it.

Undersignal scores the construction of a single submitted text, not the entity that produced it. Naming the outlet or author is attribution for the text you analyzed, labeling what was scored rather than claiming that the named entity lies, deceives, or manipulates.

Does Undersignal publish or list the reports it generates? +

No. Undersignal does not operate a public directory of reports and does not auto-generate, index, or publish reports to a public surface on its own initiative. Reports are produced only on user request, about content the user submits.

Can a low-scoring article still be misleading? +

Yes. Undersignal scores rhetorical and structural mechanics — not factual accuracy. A piece can be factually false and score 1.5 (no manipulation mechanics deployed — it simply asserts false claims without persuasive architecture). A piece can be factually accurate and score 8.0 (heavy manipulation mechanics deployed around true information).

We score how content is constructed, not whether its claims are true.

What does a score of 5.0 mean? +

It depends on the content type. A 5.0 in a political op-ed is unremarkable — advocacy framing is that format's job. A 5.0 in a corporate earnings call warrants scrutiny. A 5.0 in an academic paper is unusual. Use the construction label and intent verdict alongside the raw score.

Is a high score an accusation of intent? +

No. A high score means manipulation mechanics are present and intensely deployed. It does not mean:

  • The author intended to deceive
  • The claims are false
  • The content is harmful or illegal
  • The outlet acted in bad faith

Intent is assessed separately in the intent verdict section of each report. A declared political advertisement can score 8.5 — that is a description of its mechanics, not an accusation.

Why can't Undersignal score videos, podcasts, or audio? +

We score text mechanics. Audio and video content can be submitted as a transcript — paste the transcript directly or use a transcription service first. The same rubric applies.

Technical Reference

Model Fixed, versioned configuration; the exact model is named in every report's reproducibility certificate
Scoring Multiple independent scoring passes per report at the model's lowest-variance setting; the median result feeds the published score
Rubric version v1.0
Dimensions 14 (13 scored 1 to 10, plus Disguise Factor on 0 to 3 scale)
Clusters 6
Score range 1.0–10.0 (linear normalization, full range used)
Cache Permanent, keyed on content hash + rubric version + model ID

Questions about the methodology, calibration approach, or research applications: hello@undersignal.ai

Run a free report →