Training Model Detection Generative AI Analysis Open Web Monitoring
The foundation of the forensic record. Training data evidence feeds directly into output analysis and web monitoring. One continuous chain from model training through public deployment.
The Problem

Courts are dismissing cases for lack of evidence.

In June 2025, two federal judges sided with AI companies. Neither ruling found that training was legal. Both found that plaintiffs hadn't proven their specific works were in the data.

One judge noted that plaintiffs with better evidence will often win. That's what we're building.

Current Litigation Evidence Gap
What courts asked for
Evidence of specific works in training data with documented market impact
What plaintiffs provided
Side-by-side screenshots and general claims about publicly known training datasets
What VN delivers
Multi-layered forensic evidence packages with documented methodology, confidence intervals, and reproducible results
The Methodology

Four independent analyses.
One forensic package.

No single approach produces courtroom-grade evidence on its own. We layer four methods to build the strongest possible case.

Output Elicitation Analysis

We prompt AI models to reproduce your works, document every output, and score the similarity. A reproducible catalog of infringing generations.

  • Proprietary elicitation protocols designed around each asset profile
  • Perceptual and semantic scoring against source material
  • Full audit trail of prompts, parameters, and outputs
  • Works across all major generative AI platforms
Sample Output Elicitation Scan
Midjourney
87% match
DALL·E 3
62% match
Stable Diff.
91% match
Imagen 3
34% match
Flux Pro
71% match
Illustrative example. Actual results vary by asset and model.

Black-Box Membership Inference

Works on any model's public API with no internal access required. We measure whether a model reconstructs your content with unusual fidelity, then quantify the probability it was trained on your work.

  • Works on closed-source models (Midjourney, DALL·E, Imagen)
  • Statistical confidence intervals on every assessment
  • Purpose-built statistical framework, not adapted from general tools
  • Designed to withstand adversarial cross-examination
Membership Inference Signal
Your asset
0.94
Control (similar)
0.41
Control (random)
0.22
Anomalous reconstruction fidelity detected. High membership probability.

Perceptual Similarity Scoring

We compare your assets to model outputs across multiple perceptual and semantic dimensions, measuring not just visual similarity but how closely the model represents your specific works versus similar concepts.

  • Multi-dimensional embedding and perceptual analysis
  • Distinguishes "learned your specific work" from "learned the concept"
  • Produces quantified similarity scores admissible as expert evidence
  • Thousands of generated outputs per asset for statistical validity
Embedding Space Analysis
High match Partial Conceptual

Checkpoint Extraction

For open-source models, we recover training data directly from model weights. This is the most legally potent form of evidence available.

  • Direct extraction from model parameters for applicable architectures
  • Applicable to Stable Diffusion, Flux, and HuggingFace models
  • Produces the strongest form of evidence: literal recovery
  • Limited to open-source models with accessible checkpoints
Checkpoint Analysis Coverage
Stable Diffusion (all versions) Full extraction
Flux (open variants) Full extraction
Midjourney Black-box methods only
DALL·E / Imagen Black-box methods only
Our Commitment

We tell you exactly what we can prove. And what we can't.

Forensic integrity matters more than marketing claims. This is how we position our findings.

What we don't claim

"We can prove with absolute certainty that your specific image was in the training set." No methodology can make this claim today with zero margin of error on large-scale foundation models.

What we deliver

The strongest available evidence, using multiple independent methodologies, with documented confidence intervals and reproducible results. More than what exists today, and more than what plaintiffs recently brought to court.

Get Started

Start with a forensic assessment.

A targeted scan of your highest-priority assets against major AI models. Delivered as a forensic report with full methodology.

Request Forensic Assessment Next: AI Detection →