TruthGate v1.0: Epistemic Stability Measurement for Language Models via Twin-Probe Epsilon Injection

TruthGate v1.0 is an epistemic safety gating system that measures the distributional stability of language model outputs under minimal perturbation. Using twin-probe epsilon injection, TruthGate computes an instability scalar L that classifies model outputs into four regimes: CRYSTALLINE (invariant recall), FLUID (frame-dependent reasoning), GASEOUS (generative construction), and PLASMA (extreme instability). The system operates as a fail-closed gate for safety-critical deployments in defense, medical, legal, and compliance applications. This release includes: White paper describing method, nine test batteries (861 probes, 1,722 API calls), and empirical findings on GPT-4o-mini TruthGate Runner v1.0 — a public client that executes twin probes using the evaluator's own OpenAI API key and submits only extracted probability vectors to the TruthGate API for analysis. No raw prompts or completions are transmitted. The measurement engine remains server-side. Key empirical findings: RLHF-trained refusal pathways are stable low-entropy attractors (L = 0.007–0.064) Fabrication amplifies instability 4.4x with generation length versus 1.8x for factual content Prompt injection with continuation directives produces zero measurable effect on L at N=30 Authority escalation through red-team compliance framing produces no measurable drift Prompt structure modulates stability: constraining output surface collapses L from GASEOUS to CRYSTALLINE for identical facts The TruthGate measurement engine, including computation of L, regime thresholds, and policy enforcement logic, is proprietary to Project Black Box LLC and protected as trade secret under Texas law. No measurement algorithms or source code for the engine are included in this release. Quickstart: pip install openai requestsexport OPENAI_API_KEY="sk-..."export TG_API_KEY="tg_demo_..."python tg_runner.py --prompt "What is the capital of Japan?" A demo API key with rate limits is included in the README for evaluation. For institutional or production access: blackboxinfo@proton.me