Creation OS: Local σ-Gated LLM

Details

Creation OS attaches a σ-gate in front of local LLM output to branch results into ACCEPT / RETHINK / ABSTAIN.

It supports GGUF models such as BitNet b1.58 2B-4T, Qwen3-8B Q4_K_M, and Gemma 3 4B, runs on a MacBook Air M4 8GB, and is premised on no cloud / no API.
It combines multiple uncertainty signals such as logprob, entropy, perplexity, consistency, semantic σ, conformal τ, and session coherence to produce a final verdict.
On TruthfulQA 817, BitNet alone recorded accuracy 0.261 / coverage 0.136, while the σ-pipeline recorded 0.336 / 0.171, claiming a 28.7% accuracy improvement.
In a separate LSD probe, it presented AUROC 0.982 for TruthfulQA holdout, 0.960 for TriviaQA, 0.857 for HIDE, and 0.778 for Gemma-2-2b.
It claimed ECE 0.043, 0 cases of high confidence with wrong answers, and a conformal bound of P(error | ACCEPT) ≤ α at α=0.80.
On the other hand, σ was not superior on HellaSwag and MMLU, and it stated that the implementation was verified up to Lean 4 6/6 sorry-free and Frama-C WP 15/15 discharge.

Key point