safety-gates

OpenCode plugins ported from MiMo-Code. 5 packages: runtime, cognition, memory, safety, utilities. MIT, Bun-powered.

Reference implementation of Loop Engineering for Claude Code. 13-phase workflow, OS-level safety= gates, convergence-driven termination.

Retinal and ocular longevity controller benchmarks for safety-first partial-reprogramming and rejuvenation evaluation.

Architecture-complete green-hydrogen validation repo: safety-gated bench testing, gas verification, carbon/source accounting, BOMs, assembly walkthroughs, Gankyil-A A/B tests, and evidence receipts.

Governed agency-grounding runtime for IX-CognitionKernel Wave 6: binds intent, interpretation, permission, bounded action, feedback, memory update, replay evidence, BlackFox handoff, and human authority without AGI overclaim.

The local AI workforce control plane. Claude writes. Codex reviews. Safety gates decide.

Source-available, measurement-first pulsed-energy testbed for tri-sector storage/discharge control, derated storage, phase authority, sensor-truth checks, energy accounting, kill criteria, evidence bundles, and human-reviewed scale-up gates.

Autonomous closed-loop driver for OpenCode. 11-phase state machine, 8 safety gates, convergence-driven routing—single zero-dep binary via Bun. Part of the Loop Engineering family.

Evidence-backed quality control for LLM outputs: rubrics, claim grounding, safety gates, human review, and reports.

B2B supplier outreach kit with a four-gate pre-send safety stack: MX validation, fact-check linting, reputation caps, and SQLite suppression memory.

Auditable MCP tools for PyLabRobot: run and sample context, typed provenance, simulation by default, and explicit hardware permission.

Safety-first, research-only Deriv/MT5 EA lab: MQL5 architecture patterns, ML pipeline scaffolding, dry-run validation gates, and secret hygiene. No live trading, no real data, no models, no profitability claims.

Evidence and permission for autonomous labs: expert QC gates, workcell orchestration, provenance, recovery, and honest hardware boundaries.

Local auditable decision ledger for the Q-One quantified-life project

STARFLOW is a documentation-first execution kernel for Codex-driven software work. Define how requests become structured runs with deterministic routing, safety gates, work artifacts, and evidence-based completion.

Capture, decode, and guarded-replay tooling for no-API lab instruments; incomplete maps remain blocked from actuation.

LLM Telemetry Reasoning & PowerShell Agent Engine with Safety Gates

Building Clair, the intelligence layer for autonomous labs.

반려견 상담 오케스트레이터 — 근거 RAG와 피부 스크리닝 모델을 MCP 서브에이전트로 두고, 모델이 말해도 되는 것을 코드가 정한다 (MCP · tool calling · 결정론적 안전 게이트)

Synthetic fail-closed safety-gate demo for paper/sim research infrastructure; no execution UI.