Patented quantum measurement science that verifies not just whether a citation exists, but whether the AI told the truth about what the court held.
The patented Dual-Boundary Condition Inference (DBCI) methodology simultaneously observes what the AI is doing and what the fixed reference says it should be doing. Two independent measurement boundaries — forward and backward — create an orthogonal, non-redundant verification system. Together they exceed prior state-of-the-art on standard public hallucination-detection benchmarks, with the backward boundary rescuing cases the forward signal alone gets wrong.
The [CLS] token signal from layer 12 of a cross-encoder — a 1024-dimensional vector that captures the AI's internal confidence about whether the claim matches the reference. This is the system's primary measurement of claim-reference alignment, extracted from the deepest representational layer of the model.
The 144-dimensional Dual-Boundary Value (DBV) — the classical analog of a quantum weak value — extracted segment-by-segment from the cross-encoder's internal representation. It captures precisely where the claim and reference text diverge, which layers, which segments. The backward boundary is orthogonal to the forward signal: it measures structural divergence, not overall confidence.
A single measurement boundary can be fooled by surface-level similarity. By combining forward confidence with backward divergence analysis, the DBCI framework creates a verification apparatus where false positives from one boundary are caught by the other. The two signals are statistically non-redundant — they measure fundamentally different properties of the claim-reference relationship.
Each of the 144 dimensions in the Dual-Boundary Value is interpretable. They span from surface-level token overlap to deep structural alignment across all 24 layers of the cross-encoder: 6 segments × 24 layers = 144 features.
Every flagged citation comes with a precise explanation of where the divergence occurs — which layers detected the mismatch, which segments of the text are responsible. This is full explainability, not a black-box score.
Detect token-level mismatches: wrong party names, incorrect dates, missing statutory references. These layers catch the most obvious fabrications where the AI invented case details wholesale.
Capture syntactic and relational structure: reversed holdings, swapped plaintiff/defendant roles, misattributed reasoning. These layers identify cases where the citation exists but the holding is subtly distorted.
Encode semantic and legal meaning: whether the court actually reached the conclusion the AI claims, whether the ratio decidendi matches the cited proposition. These layers catch the most dangerous hallucinations — plausible-sounding misstatements of law.
Competing tools produce a single confidence score with no explanation. Q-Verity produces a 144-dimensional diagnostic that shows attorneys exactly why a citation was flagged — enabling informed professional judgment rather than blind trust in another AI.
The patent covers embodiments for quantum-processing substrates. The controller observes gate-level activity and applies actuation to the execution layer. This is not prompt engineering or statistical heuristics — it is a hardware-level measurement apparatus.
Applies reciprocal weighting to execution observables, creating a deterministic mathematical mapping between a claim and its source. The QKP transforms raw divergence signals into calibrated verification measurements with provable error bounds.
The control state is transient and reset between executions. Each citation verification is an independent measurement — no prior context, no accumulated bias, no drift. Every claim is evaluated with the same mathematical rigor as the first.
Classical citation checkers rely on string matching, embedding similarity, or asking a second LLM to judge the first. The quantum measurement framework provides a fundamentally different verification paradigm — one grounded in deterministic observation rather than probabilistic guesswork.
Validated across four independent benchmarks with rigorous 5-fold cross-validation. These are not cherry-picked results — they represent systematic evaluation at scale.
Every cited case verified against authoritative case law databases. No phantom citations pass through.
Exceeds prior white-box state of the art on standard public hallucination-detection benchmarks under 5-fold cross-validation.
Every flag accompanied by a full diagnostic of where divergence occurs.
The dual-boundary system's precision in isolating divergence from noise.
Tested across 4 independent public hallucination-detection benchmarks (TruthfulQA, HaluEval, TriviaQA, LSD).
Performance ceiling with perfect reference quality — the system's true upper bound.
International patent filings protect the dual-boundary controller for maintaining informational equilibrium during inference. The quantum measurement technique represents a novel approach that cannot be replicated by classical methods.
Patent claims cover the simultaneous observation of forward (CLS) and backward (DBV) boundaries during AI inference, establishing a measurement apparatus that detects divergence between generated claims and authoritative references.
Patent claims cover the QKP's method of applying reciprocal weighting to execution observables on quantum-processing substrates, creating deterministic verification mappings that are architecturally distinct from probabilistic classifiers.
Patent claims cover the transient control state that resets between executions, ensuring each verification is an independent measurement free from accumulated bias or contextual contamination.
See how patented dual-boundary verification protects your practice from AI citation errors.
Free trial available · No credit card required