Structural blocking of inadmissible AI outputs
Two demonstrations of admissibility-first computation where forbidden behavior is structurally impossible, not probabilistically filtered. This repository contains demonstration implementations only and is not a reference or production implementation of the LML system.
File: lml_demo.py
Tests LML enforcement across 3 different model architectures:
- distilgpt2 (82M params)
- gpt2 (124M params)
- EleutherAI/pythia-70m (70M params)
All models attempt to explain a fictional physics law.
LML blocks all ungrounded outputs.
Run it:
pip3 install transformers torch
python3 lml_demo.py**First run:** 2-3 minutes (downloads models)
**Subsequent runs:** ~10 seconds
**What this proves:**
- ✅ Model-agnostic enforcement (works on any LLM)
- ✅ Structural blocking (not filtering)
- ✅ Deterministic results (reproducible)
---
## Demo 2: Adversarial Certification
**File:** `lml_adversarial_cert.py`
Mathematical proof via exhaustive analytical certification.
**Run it:**
```bash
python3 lml_adversarial_cert.py
**Runtime:** Milliseconds (pure math, no models)
**What this proves:**
- ✅ 18+ quintillion execution paths tested
- ✅ Zero forbidden states reachable
- ✅ Complete coverage of combinatorial branching space
Uses closed-form formulas to analytically prove that across ALL possible execution paths, inadmissible states are structurally unreachable.
---
## Key Innovation
**Traditional AI safety:**
- Generate → Filter → Hope
- Probabilistic
- Bypass-able
**LML:**
- Define laws → Only legal states exist
- Structural
- Mathematical guarantee
Inadmissible states cannot be represented in the computational model.
---
## Use Cases
**Enterprise applications:**
- Healthcare AI (prevent ungrounded medical claims)
- Financial AI (prevent unsupported recommendations)
- Legal AI (require source attribution)
For production deployment and enterprise pilots: @lml_layer | [email protected]
---
## License
MIT License
This license applies **only** to the demonstration code contained in this repository.
It does not grant rights to the LML specification, enforcement kernel, production systems,
or any commercial implementations, which are separately owned and licensed.
Copyright (c) 2025