← All tools

full-spectrum-lab/full-spectrum-engine

Popularity 60 Updated Development & Build

Local-first governance runtime for reproducible AI inspection, RiskVector, Runestone, AuditTrace, and governance-chain generation.

githubauto-collected

Installation

A directly usable install command is not verified yet. Check the project documentation or releases.

Full Spectrum Engine

English · 简体中文

Local-first governance runtime for reproducible AI inspection, risk visibility, audit trace, and explainable simulation.

Full Spectrum Engine is the runnable engine layer in the Full Spectrum ecosystem. It turns AI behavior, knowledge-source conflicts, structured metrics, and synthetic scenarios into reproducible outputs such as RiskVector, Runestone, local audit records, and optional ESS-lite path suggestions.

Release truth

| Line | GitHub status | Meaning | | --- | --- | --- | | v1.4.0 | Latest stable release | Replay and audit hardening baseline. | | v1.5.0 | Pre-release | Enterprise-pilot candidate; additive controls over v1.4. | | v1.0.0 | Stable historical baseline | First stable local-first contract release. |

v1.5.0 already exists on both public release tracks. Its pre-release label is intentional and must not be interpreted as a production-complete enterprise platform.

This repository is for local internal validation first:

  • runnable on a local machine
  • reproducible with fixed seeds
  • auditable at sample level
  • explainable through structured outputs
  • usable without joining an external protocol network

It does not upload enterprise data by default, does not execute final business actions on behalf of an enterprise, and does not claim production governance completeness.

中文一句话

这是全频谱体系里的企业内部治理引擎:v1.5 在本地可复现计算、Profile、回放与审计基础上,增加受控企业试点所需的配置/秘密分离、最小 RBAC、脱敏、人工复核、韧性、可观测性、备份回滚和默认关闭写回的 Connector 契约。

v1.5 enterprise pilot candidate

v1.5.0 is an additive enterprise-pilot release over the sealed v1.4 replay/audit baseline:

  • configuration contains secret references, not embedded credentials;
  • authentication is limited to local operator/service reference tokens and never treats an ObservedSubject as a login identity;
  • field minimization, masking, hashing and tokenization support controlled trials;
  • human review records bind to real v1.4 evaluation events and remain append-only;
  • retry, observable timeout, health, metrics, backup and rollback primitives are included;
  • four Connector output contracts are available with writeback disabled by default;
  • protected engine directories remain byte-identical to the frozen v1.2 baseline.

Release evidence: 285 pytest cases plus 3 subtests, v1.5 black box 13/13, v1.4 regression black box 9/9, red-line verification 9/9, and zero-byte protected-core diff. This is a controlled pilot candidate, not a claim of autonomous business execution or production-complete enterprise governance.

---