ai-oversight

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

Human Infrastructure Layer — the global infrastructure that lets AI agents call human capabilities as a cloud API.

Can we detect a deceptive AI agent? A scalable-oversight study across 3 deception types and 3 detector conditions.

RL environment + GRPO-trained overseer that detects hallucination propagation in multi-agent LLM fleets 5-signal composite reward, 4 task tiers, 112 tests

Lightweight proof-of-concept for oversight-centered metrology in coding agents: workflow-aware evaluation, interrupt channels, and claim-margin reporting beyond raw success scores.

Telemetry-grounded, calibrated failure attribution for agent oversight (OTel GenAI + Who&When).

Sentinel — embeddable inline AI oversight. Wraps the host AI's output in place so the human expert can validate, correct, and audit without leaving their workflow.

Reflex-interruption path verifier for AI actions — maps risk events to tactile signal classes and races body latency against commit delay