shedu
NabuHarnessBusiness SuiteGameResearch
Demo

/ AI OVERSIGHT

/ EVIDENCE INTEGRITY

/ GOVERNED EXECUTION

Evidence,
not assumptions.

Deep dives into how AI agents should be controlled, verified, and held accountable — run as experiments, published with the cases and the code.

Read the ResearchSee It in Practice

Published Work

Can an agent
certify its own work?

01
STUDYJuly 2026

Evidence Integrity Under a Hostile Executor

A 100-case empirical study testing whether AI agents can certify their own work. Trusting an agent's self-report — or checking only that its evidence agrees with itself — approved every false claim tested, 40 out of 40. What actually worked was an independent anchor tied to the specific claim.

Read the study
02
INTERACTIVETry it yourself

Evidence Integrity Demo

Read a real evidence bundle from the study. Decide whether you believe it. Then see the ground truth — and how each review method scored it.

Try the demo

This research runs the products.

What these studies found is what the governed harness enforces — independent review, tied to the claim, with a human holding the commit.

The Governed Harness