Evaluating Bounded Autonomy in Regulated Agentic AI: A Diagnostic Harness with Constitutional Rewards, Escalation Labels, and Runtime Governance
RegLLM measures escalation, grounding, and unsafe actions for regulated agents, but only at smoke scale.
RegLLM is a diagnostic harness for bounded autonomy in regulated agentic workflows, scoring citation validity, source grounding, schema compliance, escalation correctness, constitutional alignment, and unsafe-action rate. A deterministic runtime supervisor blocks ungrounded answers and forces escalation, while the same domain constitution feeds evaluation, training rewards, and serving guardrails. An offline run of 12 examples lifts escalation recall from 0 to 0.67 and reduces the unsafe-action rate from 0.33 to 0.08 when governance is enabled. Two single-GPU Qwen2.5-3B LoRA/DPO pilots of eight examples show large configuration variance, and the authors say they do not establish reliable adapter effects or production readiness.
- Tracks grounding, schema, escalation, constitution, and unsafe actions
- Governance raises escalation recall from 0 to 0.67 on n=12
- Unsafe-action rate falls from 0.33 to 0.08 with supervision
- Qwen2.5-3B pilots vary widely and are not production evidence
From the original article
1 min read at huggingface.co
We propose RegLLM, a diagnostic harness for bounded autonomy in regulated agentic workflows. It instruments six trustworthiness signals: citation validity, source grounding, schema compliance, escalation correctness, constitutional alignment, and unsafe-action rate. Signals are distinguished by their source of supervision: programmatic verifiers, task-level escalation labels, or AI-judge scores.
By Dipankar Sarkar · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.