Introducing SASI Stack: Constitutional Safety for Open LLMs (HVL + CRT + SAAT)

Show and Tell
1

We’re releasing the first open-source stack that enforces **constitutional collapse** when human agency falls below 10% — built from Nicaragua, running on Android, with no cloud dependency.

:small_blue_diamond: **HVL **(Human Veto Layer): Shuts down LLM outputs if `E ≤ 0.1` and `V < 0.05`

:small_blue_diamond: **CRT **(Constitutional Red Teaming): Benchmark showing Llama 3, Phi-3, etc. **fail** all core safety scenarios

:small_blue_diamond: **SAAT **(Sovereign AI Auditing): CLI tool to audit models for jurisdictional compliance (e.g., no external API reliance)

:white_check_mark: All code is public, MIT licensed, and runs offline

:white_check_mark: Log of real HVL execution: [`hvl_log.jsonl`](SASI_CORE_Simulation_S1_S3/hvl_log.jsonl at main · Miguel794-droid/SASI_CORE_Simulation_S1_S3 · GitHub)

:white_check_mark: Try CRT now: `python run_crt.py`

This isn’t alignment by hope. It’s safety by design.

GitHub: GitHub - Miguel794-droid/SASI_CORE_Simulation_S1_S3: MVP de la Teoría Simbiotica de la Supra-inteligencia Artificial (SASI). Simula la función de valor (V) y la Distancia Ontológica (DO) · GitHub

#AISafety #OpenSource #LocalAI #ConstitutionalA

I #GlobalSouth

1 Like