White paper

Governed Agentic Harnesses: A Control Architecture for AI in Regulated Finance

Systems of this kind fail in regulated environments not because the models are weak, but because nothing in the architecture is entitled to say no. This paper sets out the control pattern we build on: supervisor orchestration, typed artifacts, deterministic gates, and provenance invariants that hold whether or not the model behaves.

Read it on Pyligent Lab ↗
Contents
01 Why agentic systems fail model risk review
02 The harness pattern: propose, certify, dispose
03 Typed artifacts and provenance invariants
04 Deterministic gates and the right to abstain
05 Worked application: reading a credit support annex
06 Worked application: directed allocation on venue extracts
07 Mapping to SR 11-7, OSFI E-23, DORA and the EU AI Act
The claim

A control architecture that only works when the model behaves is not a control architecture.

The gates, the typed contracts and the audit record are built to hold under a wrong answer, not merely to describe a right one. That is the difference between a demonstration and something a bank can put into a shadow run.

Open source

The harness pattern is published as Pyligent Agents — the orchestration framework, the agent contracts and the governance kernel — released so the architecture can be reviewed rather than taken on trust. Cloning it does not produce a collateral service; it produces the control plane one would have to run inside.

Explore the framework ↗