Verification Layer
High-level architecture of MTITAN's hallucination detection engine and cryptographic proof chain system. Authored by Aditya Yadav at Millimo Inc.
Ingestion & Schema Validation
Incoming payloads validated against registered schema. Structural anomalies — missing fields, type mismatches, out-of-range values — flagged immediately before entering the verification pipeline. Prevents malformed data from consuming verification compute.
Source Cross-Reference
Each data point cross-referenced against the verified source-of-truth ledger. The system checks the payload value matches (within tolerance) the canonical value stored in the MTITAN ledger for that data source and field. Mismatches are flagged as hallucination signals.
Multi-Pass Hallucination Detection
Statistical anomaly detection, semantic consistency checks, and historical pattern matching applied in parallel. Signals classified by severity:
Merkle Proof Generation
Verified data points inserted into a Merkle tree. Root hash is a cryptographic commitment to the entire verified dataset. Proofs independently verifiable by any third party without accessing the full dataset. SHA-256 Merkle proofs with compact proof format.
Audit Trail Commit
Verification result, proof hash, hallucination signals (if any), and full request metadata committed to immutable audit log. Satisfies SOX, GDPR, and emerging AI regulation traceability requirements.
Read the Full Specification
The complete verification layer specification with message formats, detection algorithm details, Merkle tree construction, and formal consistency proofs is available in the Millimo documentation.