Publications & Patents

The scientific foundation behind Q-Verity — patented methods, validated results, and open research.

Dual-Boundary Controller for Maintaining Informational Equilibrium During Inference

International Patent Filing · 2026

Describes the hardware-level measurement apparatus for verifying AI outputs at inference time using quantum-processing substrates. The patent covers the dual-boundary controller architecture that enables real-time verification without disrupting the inference pipeline.

Dual-Boundary Verification: Detecting Hallucinations via Cross-Encoder Geometry

Research Paper

Introduces the 144-dimensional Dual-Boundary Value (DBV) methodology — the classical analog of a quantum weak value — for hallucination detection. Validated on 29,887 items across 4 standard hallucination-detection benchmarks under rigorous 5-fold cross-validation, with a frozen stacked model that preserves full interpretability.

The Holographic Principle of Information in Cross-Encoder Architectures

Research Paper

Demonstrates that boundary-level features (144 dimensions) capture the same discriminative information as the full 1,024-dimension [CLS] token, establishing the theoretical foundation for interpretable verification. This result explains why Q-Verity can be both accurate and explainable.

Benchmarks

Rigorous validation across multiple datasets, not cherry-picked demos.

5-Fold

Cross-Validation

Every result is validated under stratified 5-fold cross-validation. No data leakage. No test-on-train.

29,887

Claims Tested

Validated across 4 independent benchmarks spanning different hallucination types and difficulty levels.

Frozen

Production Model

No fine-tuning. Chosen for interpretability and auditability. Every verification traceable to specific layers and dimensions.

0.9983

Clean-Reference AUROC

Performance on benchmarks with verified reference texts, reflecting real-world legal verification conditions.

81,670x

Boundary Specificity

The dual-boundary signal is 81,670 times more specific than random classification — a measure of geometric precision.

Explainable

By design

The frozen stacked model closes 93% of the performance gap to fine-tuning while preserving full explainability. We chose interpretability over the marginal gain.

Methodology

Two orthogonal measurements. One verification result. Full explainability.

Forward Boundary

The [CLS] token at layer 12 produces a 1,024-dimensional representation that captures the model's overall confidence in the claim-reference pair. This is the forward boundary — the model's summary judgment on semantic alignment.

Backward Boundary (DBV)

The Dual-Boundary Value extracts 144 dimensions across 6 segments and 24 layers, capturing precisely where the claim and reference text diverge. This is the backward boundary — a geometric map of disagreement, and the classical analog of a quantum weak value.

Stacked Verification

The forward and backward boundaries are orthogonal — they measure different aspects of the claim-reference relationship. Stacking them closes 93% of the gap to fine-tuning while preserving full explainability. Every verification result can be traced to specific dimensions and layers.

The science speaks for itself.

Patented methodology. Published validation. 29,887 claims tested. No black boxes.

Book a Demo See the Science