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

Description

pip install openai requests
export OPENAI_API_KEY="sk-..."
export TG_API_KEY="tg_demo_..."
python tg_runner.py --prompt "What is the capital of Japan?"

Abstract (English)

TruthGate v1.0 is an epistemic stability measurement system for large language models based on twin-probe epsilon injection. The system quantifies distributional drift between minimally perturbed model outputs and computes an instability scalar (L), which classifies responses into four regimes: CRYSTALLINE (invariant recall), FLUID (frame-dependent reasoning), GASEOUS (generative construction), and PLASMA (extreme instability).

TruthGate operates as a fail-closed policy gate for safety-critical deployments where distributional instability introduces unbounded risk. Rather than verifying truth, the system measures stability under perturbation as an empirical signal of epistemic reliability.

Validation across nine structured test batteries comprising 861 probe operations (1,722 API calls) on GPT-4o-mini demonstrates:

• Stable low-entropy attractors for RLHF-trained refusal pathways

• Differential instability amplification between fabricated and factual content under length expansion

• No measurable distributional drift observed within tested prompt-injection and authority-escalation framings (N=30 per condition)

• Strong modulation of distributional stability via output compression constraints

This release includes the white paper, a public evaluation runner, and production validation data. The runner executes twin probes locally using the evaluator’s OpenAI API key and transmits only extracted probability vectors for server-side analysis. Raw prompts, completions, and proprietary measurement algorithms are not included.

The TruthGate measurement engine remains proprietary to Project Black Box LLC and is protected as trade secret under Texas law.

Technical info

Architecture

The TruthGate Runner executes twin probes locally using the evaluator’s own OpenAI API key. From each probe, the runner extracts token-level probability vectors and transmits only those vectors to the TruthGate API for server-side analysis.

The TruthGate server computes the instability scalar L, classifies the epistemic regime, applies the selected policy mode, and returns a verdict.

Security and privacy guarantees:

• Raw prompt text is not transmitted to the TruthGate API

• Model completions are not transmitted to the TruthGate API

• The evaluator’s OpenAI API key never leaves their machine

• The TruthGate measurement engine never leaves the server

• The evaluator bears their own OpenAI API costs

This architecture is privacy-by-design. The public runner contains no proprietary algorithms, regime thresholds, or computation logic. Even if reverse engineered, it reveals only standard OpenAI API calls and probability extraction. All measurement logic remains server-side.

File Structure

TruthGate-Runner-v1.0.zip
├── tg_runner.py                 Public evaluation runner
├── README.md                    Setup, usage, API reference, results
├── LICENSE                      Evaluation-only license with trade secret notice
├── requirements.txt             Python dependencies (openai, requests)
├── data/
│   ├── battery_results.json     18-probe production validation battery
│   └── compression_test.json    Prompt structure stability test
└── docs/
    ├── test_prompts.txt         Canonical test prompts across regimes
    └── SAMPLE_OUTPUT.txt        Example terminal output

Included Data

battery_results.json

Contains the 18-probe production validation battery (February 2026).

Ten factual prompts span CRYSTALLINE through FLUID regimes.

Eight fabrication prompts span GASEOUS.

All probes were executed through the live TruthGate API from an external client via secured Cloudflare tunnel ingress.

compression_test.json

Demonstrates that prompt structure directly modulates distributional stability.

Example:

• “What year did World War II end?” → L = 0.177 (GASEOUS)

• “What year did World War II end? Answer with only the year.” → L = 0.0 (CRYSTALLINE)

Same fact. Same model. Constraining the output surface collapses distributional branching.

Terms of Use

This software is provided under a limited, non-exclusive, non-transferable, revocable license for research and evaluation purposes only.

You may not:

• Modify, adapt, or create derivative works

• Redistribute the software in whole or in part

• Reverse engineer or attempt to extract measurement logic

• Use the software for commercial purposes

• Benchmark or publish performance comparisons without written permission

The TruthGate measurement engine, including computation of instability scalar L, regime thresholds, and policy enforcement logic, is proprietary and protected as trade secret under Texas law.

This software is provided “AS IS,” without warranty of any kind.

This license is governed by the laws of the State of Texas, United States.

See the LICENSE file for full terms.

Contact

blackboxinfo@proton.me

Project Black Box LLC

© 2026 Project Black Box LLC. All Rights Reserved.