How Truth Engine Works
Truth Engine evaluates digital assets across three independent axes of verification, unified by an explainable AI layer. Each axis runs as an isolated forensic pipeline and is synthesized into a transparent human-readable verdict.
Defense in Depth
No single detector is enough. By cross-referencing content forensics, semantic alignment, and source intelligence, Truth Engine catches threats that isolated detectors often miss.
Axis I: Content Authenticity
Integrity Clash Auditor
Joint evaluation of cryptographic provenance (C2PA/EXIF metadata) against deep AI detection. Flags epistemic contradictions when metadata claims conflict with detection results.
FakeParts Spatial Localization
Moves beyond binary real/fake classification to generate spatial heatmaps highlighting exact manipulation regions using confidence-weighted attention maps.
ImBD Stylometric Curve Alignment
Detects machine-revised text using the Imitate Before Detect paradigm, measuring style-conditional probability curvature against AI token distribution preferences.
Axis II: Contextual Consistency
Generate-then-Detect Context Auditing
Uses Vision-Language Models to independently describe the image, then computes cosine similarity between VLM description and user claims. Low similarity triggers a semantic mismatch alert.
Axis III: Source Credibility
Automated OSINT Threat Matrix
Aggregates WHOIS domain data, VirusTotal threat intelligence, and Google Safe Browsing results to produce a cybersecurity-grade source credibility assessment.
Glass Box CoT Report
Chain-of-Thought Forensic Synthesis
Uses Llama 3.3 70B with strict CoT prompting to synthesize all axis results into a readable forensic explanation, citing specific numerical evidence from each axis.
All Free-Tier APIs
Truth Engine runs entirely on free cloud APIs: Cloudflare Workers AI (Llama 3.3 70B, Vision 11B, BGE embeddings), HuggingFace Inference API, VirusTotal, and Google Safe Browsing. Zero cost, full capability.