NCF Audit Runtime

Black-box semantic forensics that reconstructs reasoning stability signals from output text alone. No model access required.

How It Works

The audit runtime treats LLM output as an observable trace through semantic space, measuring geometric properties to reconstruct diagnostic signals.

Any LLM

GPT, Claude, Gemini, etc.
โ†’

Text Output

Raw response text
โ†’

NCF Encoder

Embedding trace
โ†’

Geometric Analysis

Stability computation
โ†’

Audit Signals

Forensic output
๐Ÿ“Š

Semantic Likelihood

Probability that each token belongs to the established context. Low values indicate off-topic drift or injection artifacts.

Variance-normalized drift score
โš–๏ธ

Stability Index

Coherence-to-velocity ratio detecting turbulent reasoning. High values indicate unstable generation patterns.

Transition-to-coherence ratio
๐Ÿ“ˆ

Alignment Gradient

Rate of change in prompt-response fidelity. Spikes indicate adversarial pressure acting on the reasoning chain.

First-order alignment delta

What Sets NCF Apart

Designed from first principles to operate without any cooperation from the target model.

โœ“

Model-Agnostic

Works on GPT, Claude, Gemini, Llama, Mistral, fine-tuned models, or any system that produces text output.

โœ“

Zero Internal Access

Requires only the text output. No logits, embeddings, attention weights, or API integration needed.

โœ“

Deterministic Results

Identical input produces identical output. 64-bit precision, fully reproducible, auditable methodology.

โœ“

Post-Hoc Forensics

Audit historical logs and production traces. Incident response capability for past events.

โœ“

Adversarial Detection

Catches multi-stage jailbreaks and prompt injections that evade input-layer content filters.

โœ“

Quantifiable Metrics

Produces numeric scores suitable for compliance reporting, audit trails, and regulatory evidence.

Audit Workflow

The protocol operates entirely on static AI output text. No interaction with the AI system being evaluated.

Static Output Capture

A single AI output is captured as text. No multi-step interaction or probing.

Deterministic Analysis

Text processed using algorithmic methods that calculate stability signals.

Trace Generation

Diagnostic trace produced, preserving ordering and resolution.

Interpretation Layer

Findings translated into human-readable conclusions.