TruthGuard
See Through the Spin
Powered by a 900M-parameter DeBERTa-v3-large model with five simultaneous task heads, TruthGuard classifies 56 span-level manipulation techniques, scores emotional intensity per token, detects 7 emotion classes, and runs 8 stealth propaganda analyzers on every article you read. It also rates content safety across 4 audience tiers — from COPPA-compliant for kids to relaxed for adults — so you can protect your family while seeing through the spin.
Challenges You Face
Invisible Manipulation
Articles that are 100% factually accurate but use selective sourcing, emotional quote stacking, and strategic hedging to push an agenda
Social Media Overwhelm
Can't tell genuine content from manufactured outrage, engagement bait, and coordinated messaging campaigns
Emotional Exploitation
Content designed to trigger fear, anger, or FOMO — detectable via our 7-class emotion classifier running per-token
Child Safety Online
Kids exposed to harmful content, predatory behavior, and age-inappropriate material across platforms
Hidden Bias in News
Stealth propaganda that passes every fact-check: one-sided sourcing, context omission, passive framing
Family Arguments
Loved ones sharing manipulative content — now you can explain exactly what techniques are being used and why
Key Features
56-Technique Span Detection
Every manipulation technique highlighted in-text with confidence scores and severity ratings
Manipulation Density Score
Overall propaganda density and severity calculated from span coverage across the full article
Emotional Arc Analysis
See which of 6 fundamental story shapes an article follows — and where emotional intensity peaks
7-Class Emotion Detection
Per-span emotion classification: anger, fear, joy, sadness, disgust, surprise, or neutral
LLM-Powered Explanations
Plain English explanations of each technique, generated by Ollama, OpenAI, Gemini, or Claude
Stealth Bias Alerts
8 analyzers catch source imbalance, quote stacking, hedging, context omission, and implied narratives
Content Safety Ratings
Automatic G/PG/PG-13/R/X ratings using Detoxify toxicity scoring and NudeNet image analysis
Grooming Detection
4 sub-analyzers (intimacy escalation, secrecy patterns, age probing, stage classification) monitor chat
URL Safety Scanning
Google Safe Browsing API checks for malware, social engineering, unwanted software, and harmful apps
What Others Miss
Stealth manipulation that hides behind objectivity. Heimdall exposes it.
Measures perspective_ratio of quoted sources — fires when 80%+ of sourced perspectives favor one side (minimum 2 sources). Uses SpaCy NER + attribution verb detection.
Detects statistical claims without baselines, event claims without evidence, truncated quotes, and causal assertions without proof. Severity boosted 1.2x in lead paragraphs.
Matches 40 doubt-planting patterns across 3 severity tiers. Reduces severity 30% when properly attributed. Fires when hedging density exceeds 0.02 markers/char.
Runs VADER sentiment on each extracted quote. Detects when emotional asymmetry exceeds 0.5 — e.g., all passionate quotes from one side, dry statements from the other.
Detects rhetorical questions, temporal juxtaposition implying causation, recency effect placement, and strategic anecdote-before-statistics sequencing.
Techniques We Detect
Fear of missing out — fake urgency and scarcity pressure
Content engineered to trigger anger for engagement and sharing
"Like if you agree!" — social pressure tactics for algorithmic amplification
Emotionally charged words substituted for neutral alternatives to sway perception
False consensus via "everyone's doing it" framing
Weaponizing idealized past to discredit present or drive action
Pricing
Personal
- ✓Unlimited analyses
- ✓All 56 techniques + 8 stealth
- ✓Emotional arc analysis
- ✓LLM explanations
Family
- ✓5 accounts
- ✓Content safety ratings
- ✓Grooming detection
- ✓URL scanning
- ✓Parental alerts
Your BS Detector, Supercharged
Schedule a demo or start your free trial today.