Heimdall Intelligence
Beyond Fact-Checking: Map the Manipulation Architecture
For serious news analysts who need to understand not just what is being said, but the complete manipulation architecture behind it. Our narrative flow engine maps 16 manipulation choreography patterns across 6 fundamental story shapes (Reagan et al. 2016), while 8 stealth propaganda analyzers quantify source perspective ratios, emotional quote asymmetry, word-choice framing, hedging density, and context omission — the manipulation that passes every fact-check.
Challenges You Face
Manipulation Choreography
Articles sequence techniques deliberately — sympathetic hooks, fear escalation, call-to-action timing — and you need to see the choreography
Stealth Framing
Factually accurate articles that manipulate through selective sourcing, quote stacking, and passive voice patterns
Source Comparison
Same story across outlets — different emotional arcs, different sourcing ratios, different implied narratives
Headline Manipulation
Headlines that sensationalize, contradict the body, or use clickbait patterns to drive engagement
Revision Tracking
Articles change after publication — headlines shift, context is added or removed, framing evolves
Echo Chamber Quantification
Need data, not intuition, to prove systemic bias patterns across outlets
Key Features
16 Manipulation Flow Patterns
Detect fear escalation, sympathetic hooks, fear-to-solution, anger buildup, technique accumulation, confirmation bias exploitation, anchoring, firehose of falsehood, and more
6-Arc Emotional Trajectory
Classify narrative shape: Rags to Riches, Riches to Rags, Man in a Hole, Icarus, Cinderella, Oedipus — with confidence scoring via Pearson correlation
Freytag Structure Mapping
Automatic detection of Exposition, Rising Action, Climax, Falling Action, and Resolution stages — refined around detected emotional peak
Headline Accuracy Analysis
POST /analyze/title returns accuracy_score, sensationalism_score, verdict (accurate → clickbait), and specific issues like question_bait and emotional_amplifier
Revision Comparison
POST /analyze/revision compares article versions — diff_summary, change_classifications, editorial_intent, and narrative flow shift analysis
8-Analyzer Stealth Dashboard
Source balance (perspective_ratio), quote sentiment (VADER), framing (SpaCy NER + valence), voice asymmetry (dependency parsing), hedging (40 patterns), context gaps, implied narrative
Per-Span Severity Scoring
Each detected span scored on confidence, intensity (regression head), and emotional charge (emotion head) — aggregated into overall severity
LLM-Augmented Explanations
Optional per-technique explanations with counter-arguments, generated via Ollama (default: llama3.1:8b), OpenAI (gpt-4o-mini), Gemini, or Claude
Signal Intelligence
Recently shipped capabilities powering the Heimdall pipeline.
Signal-to-Noise Ratio
Measures information quality versus rhetorical noise across 6 sub-metrics. Composite score 0-100 with noise bands. Does not judge truth — measures how much of the content is signal versus noise.
Content Word Ratio
Loaded Language
Claim-Evidence Ratio
Hedge Density
Repetition
Semantic Density
Noise Bands: Low (0-33), Moderate (34-66), High (67-100). User signals include loaded_language, low_evidence, deniability_language, circular_rhetoric, and low_substance.
At a Glance
Powered by local Gemma 4 model — on-premises, no cloud API costs. Generates structured cliff-notes summaries replacing RSS descriptions with distilled intelligence.
Summary
Distilled article summary
W5H
Who / What / When / Where / Why / How
Context
Background and framing
Claims
Key assertions extracted
Sources Cited
Referenced sources listed
Claims extraction enables future fact-checking integrations.
Readability & Writing Level
Consensus reading level computed across multiple algorithms. Writer-versus-audience level gap analysis detects when content is written below its audience's level — a manipulation risk flag indicating deliberate simplification to bypass critical thinking.
5-Pathway Influence Score
Updated from 4 pathways to 5. Final score equals the maximum across all pathways, scaled 0-100.
Overt Propaganda
Density, severity, and diversity of detected propaganda techniques
Stealth Manipulation
Stealth analysis score combined with low evidence and loaded language signals
Emotional Manipulation
Narrative flow score combined with loaded language and sentiment analysis
Balanced Manipulation
All dimensions weighted together plus noise score
Information Quality
NEWNoise score, unsupported claims, repetition, and writing level manipulation risk
Narrative Flow Analysis
6 shapes sampled at 5 reference points (0%, 25%, 50%, 75%, 100%), classified via Pearson correlation against reference templates
Peak emotional intensity position mapped onto Freytag stages, with adaptive boundaries around the detected peak
16 patterns across 4 categories: escalation (fear, anger, technique accumulation), structural (sympathetic hook, fear-to-solution), correlation (editorial, fear-without-facts), intelligence (confirmation bias, anchoring, firehose)
Combined measure of narrative absorption — high emotional intensity + low factual grounding = reduced critical thinking
Same story analyzed across sources — compare arc shapes, stealth scores, technique distributions, and implied narratives
What Others Miss
Stealth manipulation that hides behind objectivity. Heimdall exposes it.
Extracts quoted sources via 4 regex patterns + SpaCy NER for PERSON/ORG near attribution verbs. Classifies each source perspective as FAVORING_NARRATIVE or OPPOSING_NARRATIVE. Fires when perspective_ratio >= 0.8 or <= 0.2.
Extracts direct quotes (10-500 chars), runs VADER compound sentiment. Emotional threshold: |compound| >= 0.3. Detects asymmetry >= 0.5 or all emotional quotes on one side.
Uses SpaCy NER to find entities appearing 2+ times. Computes avg VADER sentiment of surrounding context per entity. Detects valence asymmetry > 0.2 between entity pairs.
Same entity extraction. Counts active vs. passive voice per entity via SpaCy dependency parsing (nsubjpass, auxpass). Detects voice asymmetry > 0.2.
19 high-doubt patterns (severity 0.5-0.9), 15 medium-doubt (0.4-0.6), 6 low-doubt (0.2-0.35). Reduces severity 30% when attributed to named source. Density: weighted_sum / text_length × 1000.
4 claim types: statistical (% without baseline), event (unprecedented without evidence), quote (truncated), causal (attribution without proof). Checks for contextualizing patterns within 300 chars.
4 methods: rhetorical questions (8 patterns), temporal juxtaposition (causation via proximity), recency effect (emotional details at paragraph end), strategic sequencing (anecdote-before-statistics).
Techniques We Detect
Misleading Headline
Title-body disconnect — sensationalism_score and accuracy_score quantify the gap
Cherry Picking
Selective facts detected via statistical claim analysis and context omission patterns
Out of Context
Quote truncation and isolation — detected at span level with surrounding context analysis
False Equivalence
"Both sides" framing when sides are not equivalent — detected via source balance asymmetry
Statistical Manipulation
Percentages without baselines, counts without timeframes, trends without evidence
Coordinated Messaging
Synchronized talking points detected across multiple articles via technique fingerprinting
Pricing
Analyst
$14.99/mo
- ✓ Full 56-technique analysis
- ✓ Stealth detection
- ✓ Flow patterns
Professional
$49.99/mo
- ✓ 500 analyses/day
- ✓ API access
- ✓ Bulk analysis
- ✓ Revision tracking
- ✓ LLM explanations
Newsroom
Custom
- ✓ Unlimited
- ✓ Team accounts
- ✓ CMS integration
- ✓ White-label
- ✓ Editorial audit tools
The Next Level After Fact-Checking
Schedule a demo or start your free trial today.