How ProofLayer Works: Under the Hood

A technical deep-dive into ProofLayer AI's multi-model detection pipeline. ## Architecture Overview ProofLayer uses a multi-provider architecture where content is analyzed using advanced AI models from OpenAI, Anthropic, and Google. Our proprietary scoring algorithm combines these analyses into a unified trust score. ## The Verification Pipeline ### 1. Content Ingestion When content is submitted via our API, it goes through: - File validation (type, size, format) - Content hashing (SHA-256) for deduplication - Cache lookup for previously analyzed identical content ### 2. Provider Selection The system selects the optimal AI provider based on: - Content type (audio → OpenAI Whisper, images → GPT-4o Vision, etc.) - User preference (BYOK provider selection) - Provider availability and response times ### 3. AI Analysis Each provider runs specialized detection prompts that analyze: - Text: Linguistic patterns, statistical distributions, AI-isms - Images: Texture artifacts, lighting consistency, hand/finger anomalies - Audio: Prosody patterns, spectral analysis, breath sounds ### 4. Trust Score Calculation Our proprietary algorithm converts raw analysis into a 0-100 trust score: - Raw AI probability is inverse-mapped to trust (100 = fully authentic) - Confidence weighting pushes uncertain results toward 50 - Provider agreement increases confidence when multiple signals align ### 5. Model Fingerprinting We maintain a database of model signatures — unique patterns left by different AI generators. This allows us to identify not just IF content is AI-generated, but WHICH model created it. ## BYOK Architecture BYOK keys are encrypted at rest using AES-256-GCM. When a scan is initiated: 1. The system checks if the user has a BYOK key for the selected provider 2. If yes, the key is decrypted in memory and used for the API call 3. Only the platform fee is charged (not counted against the standard scan quota) 4. The key is never logged or stored in plaintext ## API Performance Typical processing times: - Text analysis: 1-3 seconds - Image analysis: 2-5 seconds - Audio analysis: 3-10 seconds (includes transcription) Explore our API documentation or try it yourself.