Our Model
Metis
Verifex's document-first Trust VLM, engineered to establish trust, detect fraud, and surface issues across your document estate.
Metis is Verifex's document-first Vision-Language Model for fraud detection, tamper analysis, provenance assessment, and trust scoring across high-stakes business documents. It is designed for workflows where teams need evidence-backed decisions, not just extraction output.
Three Metis model tiers
Metis is available in three product tiers so customers can match cost, evidence depth, and detection quality to the workflow.
Metis Lite
The free preview model used for compatibility checks, basic trust scoring, and limited evidence visibility.
Best for: Quick previews before a paid rollout.
Metis Advance
The stronger paid model that exposes full evidence, deeper explanations, and better day-to-day verification quality.
Best for: Operational teams that need stronger results without enterprise deployment work.
Metis Pro
Our best model for enterprise customers, tuned for the highest accuracy, the deepest detection stack, and production deployment needs.
Best for: High-stakes enterprise verification workflows.
Built to establish trust - not just read documents
Metis is different because it was designed from day one to answer a single hard question: can we trust this document - and if not, why?
Document-first hybrid fusion architecture. Metis is built as a hybrid visual model that combines pixel-level indicators, OCR extraction, CNN-based visual encoders, and word-level anomaly detection in a fused decision stack. This allows the model to reason jointly over layout, image evidence, extracted text, and semantic inconsistencies instead of relying on a single modality.
Document-native reasoning. Metis is tuned for invoices, contracts, KYC files, government forms, and other high-risk business documents where layout, provenance, and contextual consistency matter as much as raw text extraction.
Trust & provenance tracking. Metis reasons about origin, modification history, embedded metadata and visual artifacts to produce a provenance score and explainable evidence for every verification decision.
Tamper & manipulation detection. Specialized detectors find digital edits, scanned retouching, copy-paste inconsistencies, altered line items, forged stamps/signatures and other tampering signals that typical OCR or LLM workflows miss.
Multimodal, cross-document reasoning. Metis correlates fields across pages and documents (for example invoice to purchase order to contract) to flag contradictions, missing evidence or suspicious patterns.
Core verification & fraud capabilities
Metis provides a comprehensive toolset for making documents auditable and defensible:
- Fraud & anomaly scoring. Probabilistic scoring that highlights suspicious documents and ranks issues by severity.
- Tamper forensics. Visual and metadata analysis that identifies likely edits, overlays and altered scans - with visualized evidence for each finding.
- Signature & seal verification. Multi-signal signature analytics (visual match, stroke analysis, contextual validation) plus seal/stamp authenticity checks.
- Cross-field reconciliation. Automatic cross-checks (amounts, dates, tax IDs, account numbers) across pages and related documents to detect discrepancies.
- Semantic table & line-item integrity. Table-aware extraction that preserves relationships and flags inconsistent or suspicious line items.
- Robust OCR + handwriting understanding. Industry-grade recognition across low-quality scans, handwriting and multilingual documents.
- Explainability & evidence bundles. Human-readable rationales, confidence metrics and the underlying evidence (image patches, bounding boxes, metadata) for every decision.
- Policy automation & human-in-the-loop. Declarative policy rules and escalation paths so teams can enforce compliance and continuously improve model performance.
Benchmark
Metis benchmark on the Verifex validation set
Metis is benchmarked on the Verifex dataset, a document corpus assembled from real customer documents, curated synthetic data, and open-source material. The validation benchmark below compares Lite, Advance, and Pro on the same held-out customer validation set, manually annotated for evaluation.
Training corpus
979,561
Unannotated documents used for large-scale document representation learning.
Validation split
12,464
Held-out customer documents, fully manually annotated for evaluation.
Metis Pro AUC
0.9759
Metis Pro Accuracy
91.21%
Metis Pro Precision
95.97%
Metis Pro Recall
83.22%
Metis Pro F1
89.14%
Validation benchmark
Shared validation split with fully manual annotations on customer documents. Positive class: tampered or fraudulent documents.
| Model tier | AUC | Acc | Prec | Recall | F1 |
|---|---|---|---|---|---|
| Generic OCR + LLM pipeline | 0.6101 | 0.6040 | 0.6471 | 0.5341 | 0.5852 |
| Metis Lite | 0.8147 | 0.7837 | 0.8479 | 0.7114 | 0.7737 |
| Metis Advance | 0.9174 | 0.8814 | 0.9147 | 0.7910 | 0.8484 |
| Metis Pro | 0.9759 | 0.9121 | 0.9597 | 0.8322 | 0.8914 |
Evaluation notes
Dataset composition and score interpretation.
- The Verifex dataset combines real customer documents, curated synthetic data, and open-source material.
- Positive class: documents labeled as tampered or fraudulent.
- All reported benchmark scores below are measured on the validation split only.
- The validation split is entirely customer-sourced and manually annotated to reflect real verification conditions.
Operational value beyond the benchmark
Metis was engineered for high-stakes workflows where errors cost money and reputation:
- Customer-grounded evaluation. Reported benchmark scores are measured on a validation split composed entirely of real customer documents, manually annotated for evaluation.
- Clear tier progression. Benchmark results improve consistently from Metis Lite to Metis Advance to Metis Pro on the same customer validation set.
- Precision-first operating quality. Metis Pro combines high precision and strong F1 to reduce false-positive review burden while preserving broad issue coverage in document verification workflows.
- Auditable decisions. Every alert includes provenance, confidence and visual evidence suitable for compliance reviews, dispute resolution, and human escalation.
We provide pilot evaluations and tailored benchmark summaries to quantify performance on your document set.