What's this about?
Simulated evaluation is used everywhere to benchmark agents, but nobody has systematically quantified how much evidence a simulated pass actually gives you about physical deployment. SimVerity is a verdict-transfer assurance framework that tries to answer that.
What is "verdict transfer"?
If your agent passes in simulation, that is a verdict. Verdict transfer asks how much that verdict is worth in the real world — expressed as a confidence statement rather than a yes/no. It is the difference between "it worked in the simulator" and "given how this simulator differs from reality in these specific ways, a pass here corresponds to roughly this much real-world confidence."
Why does this matter to you?
- Publishing the same day as Microduck's sim2real pipeline makes the pairing hard to miss.
- It is the discipline that safety evaluations of embodied agents currently lack — everyone reports simulator results and hopes.
- The framing generalises past robotics to any evaluation run in a mock environment, including security testing.