Research

Findings from building a control runtime for agents that take real actions. Everything here comes from running code against real catalogs, real tools and real models — the numbers are reproducible, and where a result is a single illustration rather than a benchmark, it says so.

Papers

GSAR: Typed Grounding for Hallucination Detection and Recovery in Multi-Agent LLMs — Federico A. Kamelhar, 2026. Partitions every claim an agent makes into grounded / ungrounded / contradicted / complementary against typed evidence, scores the partition, and routes the result to proceed, regenerate, replan or abstain. Evaluated with multiple LLM judges on FEVER.

Implemented in the SDK as GSAR; ground_finding() returns evidence or an abstention, so an ungrounded claim is not something the caller can accidentally ship.

Findings

The family of harm your agent policy cannot see — an agent risk policy tends to encode one family of consequence and stay silent about the others, and the blindness survives code review, its own tests, validation against the real tool catalog, and training. Three independent instances, including a measured one: Clusiana-Admit-4B, a classifier we train for exactly this job, scored 6/6 on destruction and 1/6 on standing commitment — and enumerating the missing families in the policy made overall recall worse, not better. 64-case dataset, two conditions, numbers and limits on the page.