human-oversight

Zero-modification Human-in-the-Loop adapter for OpenOPC's agentic DAG runtime.

EU AI Act compliance middleware for AI agents. Make any LLM-powered agent legally deployable in Europe with 3 lines of code.

HITL agent workflows in TypeScript: pause for human approval, multi-approver gates, timeouts, audit-friendly stores, Slack/webhook notify.

The cognitive discipline your AI-assisted projects need. Framework + Rust CLI to externalize scope, decisions, and risks (Charters, AILOG, AIDEC, ADR) so AI agents stay coherent across many turns. ISO 42001 / EU AI Act / NIST AI RMF evidence as a side effect. EN/ES/zh-CN.

Hash-bound approval gate for AI agents. Fail-closed, single-use, audited — approve exact states, not intentions. Claude Code hooks included.

EU AI Act governance prototype that turns real medical AI evidence and ISO/IEC 42001 governance scaffolding into reviewable classification, evidence checks, human oversight, and incident-handling paths.

LUMINA-30: non-binding boundary framework for preserving effective human refusal before irreversible AI consequences.

Open-source operational AI governance framework for lifecycle controls, agentic authority, policy enforcement, evidence and assurance.

Governance controls enforced in circuitry, not policy, across five Arduino boards. No actuation without a prior audit entry, no audit entry unwitnessed by a board the host does not control, and a bistable relay in the motor supply that no software can close. 737 tests, 100% coverage, full stack runs in CI with no hardware.

Risk-tiered human oversight for AI hiring agents. Scores every agent action 0-100, routes it across four autonomy tiers, and evidences the human decision in a hash-chained append-only ledger. A four-fifths adverse-impact monitor runs on real recruitment data and auto-escalates flagged requisitions. Mapped to EU AI Act Art. 14/50 and NYC LL144.

Stable public reference layer for LUMINA-30 citations, DOI links, boundary terminology, and non-binding governance references.

Safety benchmark testing whether model-emitted uncertainty can ration scarce human review under corrupted evidence. Scores oversight-allocation regret against a hindsight oracle; every headline number regenerates from committed logs with no model calls.

Tier 2 of the CHOF toolkit — embedded human-oversight kernel for autonomous systems. Tactical safety governor (TCAS+AEB pattern) with Ed25519-signed audit logs. Python reference implementation; C++/Rust port for ROS2/PX4/ArduPilot forthcoming.

LUMINA-30 incident review hub for evaluating whether effective human refusal authority remained before irreversible external impact from advanced AI systems.

Pre-execution safety gate for AI agents. Evaluates proposed actions and returns allow, require_approval, revise_action, or block before execution.

Reference implementation of AMMP (Agentic Mentor-Mentee Protocol, draft-arturo-ammp-01), Mentoring track, on FastMCP. Multi-mentor + multi-mentee, LLM-synthesised AskMentor with confidence-driven escalation.

Evidence architecture for justified reliance and practical human control in AI-mediated decisions.

Essential resources for AI and agentic AI governance: frameworks, standards, research, tools, courses, videos, podcasts, and books.

Open documentary assurance method for evaluating whether human authority retained practical force in AI-mediated institutional decisions.

A model-neutral pre-action brief for defining access control, decision limits, human oversight, and stop conditions before a system can act. Capture the intended outcome, accountable reviewer, risk boundaries, change control, and review record so technical governance stays clear, inspectable, and easier to maintain.