safety-gates

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

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

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.

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.

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

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

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.

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

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

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.

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.

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.

Building Clair, the intelligence layer for autonomous labs.

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

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.