Skip to content

SAFE-CSM (LLM Edition): SAFE-layer external observation framework for contextual stability evaluation in LLM outputs.

Notifications You must be signed in to change notification settings

jordancapri1231-prog/SAFE-CSM-LLM-Edition

Repository files navigation

SAFE-CSM — LLM Edition

SAFE-CSM (LLM Edition) is a SAFE-layer external observation framework designed to stabilize contextual drift in Large Language Model outputs.
This repository contains the public demonstration code, figures, and documentation for the LLM-focused version of SAFE-CSM.


English

Overview

SAFE-CSM (LLM Edition) provides:

  • External observation of Δφ drift in LLM contextual output
  • Lightweight stability feedback without modifying or training the model
  • Evaluation tools for comparing “LLM only” vs “LLM + SAFE-CSM”
  • Fully transparent and SAFE-compliant implementation

This edition does not include reinforcement loops, adaptive control, psychological models, or internal inspection.
All analysis is limited to externally observable output signals.

Contents

  • SAFE-CSM_LLM_demo.py
  • SAFE-CSM_LLM_demo.txt
  • SAFE-CSM_LLM_Report.txt
  • SAFE-CSM_LLM_FullDescription.txt
  • LICENSE_SAFE.txt
  • Figures
    • figure_1_fixed_to_release.png
    • figure_2_sink_control.png
    • figure_3_diverggence_stabilized.png

Related Projects

  • SAFE-CSM (Industry Edition)
  • CSM SAFE Core (Initial Evaluation Edition)

Citation (OSF DOI)

Please cite the OSF DOI for reproduction and academic use:
[10.17605/OSF.IO/VAURE]


日本語 / Japanese

概要

SAFE-CSM(LLM版)は、LLMの文脈出力における
「ドリフト・固定化・過同期(オーバーシンク)」を
外部から観測し、SAFEバンドへ戻すための
SAFE層・非適応型の観測フレームワーク です。

モデル内部には一切アクセスせず、
学習・最適化・強化なども行いません。

含まれるファイル

  • SAFE-CSM_LLM_demo.py(実行可能デモ)
  • SAFE-CSM_LLM_demo.txt(説明)
  • SAFE-CSM_LLM_Report.txt(SAFEレポート)
  • SAFE-CSM_LLM_FullDescription.txt(拡張説明)
  • LICENSE_SAFE.txt(ライセンス)
  • 図(3種)
    • figure_1_fixed_to_release.png
    • figure_2_sink_control.png
    • figure_3_diverggence_stabilized.png

関連プロジェクト

  • SAFE-CSM(産業版)
  • CSM SAFEコア(初期評価版)

引用(OSF DOI)

学術研究で使用する場合は、以下のDOIをご利用ください:
[10.17605/OSF.IO/VAURE]


Additional Notes / 追加情報

This repository provides a SAFE-layer demonstration for evaluating oversynchronization (fixation) and over-divergence in LLM-generated contextual drift, expressed as contextual phase difference (Δφ).

This repository contains only the SAFE-layer demonstration. Full operational SAFE-CSM implementations (LLM-integrated / multi-factor cognitive versions) exist separately as NDA-protected, non-public codebases.

Researchers and organizations working on AI safety, LLM stability evaluation, contextual drift mitigation, or analysis are welcome to contact us for SAFE-layer information exchange or joint evaluation.

本リポジトリは、LLM の文脈出力における 「過同期(固着)/過発散」を位相差(Δφ)として評価し、 SAFE帯域へ戻す動作を示す SAFE 層デモです。

※ 本リポジトリは SAFE 層デモのみを公開しています。 実運用向け SAFE-CSM 実装(LLM統合版/多因子版)は、 NDA 前提の非公開コードとして別途存在します。

AI安全性・安定化・文脈ドリフト解析に 関心のある研究者・企業の方は、SAFE層の範囲での 情報交換や共同検証の相談を歓迎します。

Research & technical inquiries: jordan.capri.1231@gmail.com

Main Hub: https://github.com/jordancapri1231-prog/SAFE-CSM

About

SAFE-CSM (LLM Edition): SAFE-layer external observation framework for contextual stability evaluation in LLM outputs.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages