Evaluating Bounded Autonomy in Regulated Agentic AI: A Diagnostic Harness with Constitutional Rewards, Escalation Labels, and Runtime Governance
Dipankar Sarkar
Skelf Research
Published September 28, 2026 · on the feed October 1, 2026 · ▲ 3 · arXiv 2609.37501
The abstract
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. A deterministic runtime supervisor blocks ungrounded answers and forces escalation, logging interventions. The same domain constitution informs evaluation, training rewards, and serving guardrails. Task-level should-escalate labels make the act-versus-defer decision a measurable training signal. We demonstrate the harness at smoke scale. An offline reference run (n=12) lifts escalation recall from 0 to 0.67 and reduces unsafe-action rate from 0.33 to 0.08 when governance is enabled. Two single-GPU Qwen2.5-3B LoRA/DPO pilots (n=8, same seed and evaluation split) expose substantial variation: nominally identical RL-base configurations yield task success of 0.25 versus 0.12 and escalation recall of 1.0 versus 0.5. An answer-quality adapter changes recall from 1.0 to 0.5 in Run A, but from 0.5 to 1.0 in Run B. An escalation-aware variant produces no measurable change in Run B. These small pilots do not establish reliable adapter effects or production readiness. Their contribution is diagnostic: configuration variance can overwhelm apparent tuning effects on bounded-autonomy metrics, motivating larger evaluation sets and repeated runs.
The Brief
2:28A short podcast summary of the paper. Written and voiced by AI.
Transcript
Evaluating Bounded Autonomy in Regulated Agentic AI. A Diagnostic Harness with Constitutional Rewards, Escalation Labels, and Runtime Governance. By Dipankar Sarkar, from Skelf Research. Published on the twenty-eighth of September. No code repository is linked to it.
In a regulated workflow, an agent is not meant to act freely. Its autonomy is bounded, and a recurring choice is whether to act or to defer and escalate. That decision needs to be measured, and ideally trained on, which requires treating it as a signal in its own right.
RegLLM is a diagnostic harness built for that setting. It instruments six trustworthiness signals: citation validity, source grounding, schema compliance, escalation correctness, constitutional alignment, and the rate of unsafe actions. The signals are told apart by where their supervision comes from. Some come from programmatic verifiers, some from escalation labels set at the level of the task, and some from scores given by an AI judge.
Around the measurements sits a deterministic runtime supervisor. It blocks answers that are not grounded, forces escalation, and logs every intervention. One domain constitution feeds three places at once: evaluation, training rewards, and the guardrails used when serving. Task-level labels saying when a case should be escalated turn the act-versus-defer decision into something a model can be trained against.
The harness is demonstrated at smoke scale. In an offline reference run of twelve cases, enabling governance lifts escalation recall from zero to zero point six seven, and cuts the unsafe-action rate from zero point three three to zero point zero eight.
Two small pilots follow, each on a single GPU with Qwen2.5-3B using LoRA and DPO, on eight cases with the same seed and evaluation split. They show how much results swing. Configurations of the RL base that are nominally identical give task success of zero point two five in one and zero point one two in the other, and escalation recall of one point zero against zero point five. An answer-quality adapter moves recall from one point zero down to zero point five in Run A, yet from zero point five up to one point zero in Run B. An escalation-aware variant makes no measurable difference in Run B.
The author is explicit that pilots this small do not establish reliable adapter effects, or readiness for production. The contribution claimed is a diagnosis: variance between configurations can swamp the apparent effect of tuning on bounded-autonomy metrics. Larger evaluation sets and repeated runs are the practical consequence.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
An agent that answers well is not enough. It must know when not to answer.
You deploy an AI agent to answer financial compliance questions from a regulator's rulebook. A wrong answer can mean a fine or a harmed client, so on some questions it must hand over to a qualified human.
The road ahead
- 1 · The problem ~7 min An agent answering compliance questions must know when to hand over to a human, and accuracy scores never tell you if it does.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Label each task with whether it should go to a human, sort signals by who supervises them, guard the agent at runtime, and use one constitution throughout.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min The agent runs a task, a fixed supervisor blocks or escalates and logs what it did, then six signals score the run.
- 6 · The experiment ~6 min You’ll guess: In two nominally identical GPU pilots, what did the answer-quality adapter do to escalation recall?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.