SimVerity: When Does Simulated Agent Success Survive Physical Deployment?

In the quiet corners of our homes, a new kind of intelligence is waking up. It is not a robot that walks or a machine that thinks in the way humans do, but a digital assistant that lives inside our walls, ready to turn on lights, lower blinds, or lock doors with a single spoken command. For years, engineers have tested these digital helpers in a safe, virtual world. They build a perfect copy of a house inside a computer, where every light switch and sensor behaves exactly as the code says it should. If the computer program says the agent succeeded, the engineers give it a green light to go into real homes. This process relies on a simple assumption: if it works in the simulation, it will work in reality. But as these agents move from screens to physical spaces, a critical question has remained unanswered. Does a successful test in a virtual room guarantee that the device will actually do the job in a real house, where wires are old, lights flicker, and time moves differently?

A team of researchers at Imperial College London set out to answer this by building a system they call SimVerity. Instead of trusting the computer's word, they created a method to replay the exact same scenarios in a real home while watching with a camera that acts as an independent witness. They did not just ask if the agent finished the task; they watched the entire process, second by second, to see if the physical world agreed with the virtual one. What they found was a stark disconnect. In their tests, a sophisticated computer simulation passed every single trial for turning off a light. The simulation declared the job done. Yet, when the researchers watched the real light bulb with a high-speed camera, they saw it stay on for a fraction of a second after the agent said it was off. In the virtual world, success is a static fact: the light is either on or off. In the real world, success is a process that takes time, and that time is where things go wrong.

The researchers discovered that the moment you check for success matters more than the success itself. They measured four different ways to judge if an agent had done its job: whether the software said it was done, whether the agent reported the right state, whether a human could see the effect, and whether the system had settled into a final, stable state. In their experiments, an agent could pass the first three checks and still fail the last one. In one specific test involving a light switch, the simulation said the light was off. The software reported it was off. But the camera, which served as the ultimate truth-teller, caught the light glowing for 42 separate instances where the simulation had missed the failure. These were not slow, obvious mistakes. They happened in less than a second, invisible to the standard checks that usually clear an agent for deployment. The simulation had given a false pass, approving a device that was, in that split second, still doing the wrong thing.

To understand why this happens, the team looked at how these systems are built. A simulation sees a task as a single point in time. If the command is sent and the state changes, the job is done. A real home, however, is full of delays. A signal travels through a network, a switch flips, a bulb warms up, and a sensor reports back. These steps take time. The researchers found that if you check the result too quickly, you might see the light still on, even though the agent has already told the system to turn it off. This is not a bug in the agent's intelligence; it is a mismatch between how the simulation counts time and how the physical world moves. The study showed that simply waiting a little longer before checking the result could change a success rate from 44 percent to 62 percent. The agent was not failing; the timing of the observation was the variable that decided the outcome.

The most surprising part of the research was that these failures were predictable. The team built a risk profile, a kind of map that learned from the early tests to guess where the failures would happen next. They froze this map before they started the final tests, so it could not rely on the answers. When they ran new trials, this map successfully predicted which simulations would fail in the real world, even on paths it had never physically measured. It beat a standard baseline that looked only at the device type or the general path, proving that the specific details of the task and the timing of the check were the key to spotting the danger. This means that before an agent is ever shipped to a home, engineers could use this method to say, "This looks good in the computer, but based on our map, it will likely fail in the real world at this specific moment."

The study also tested whether having two different computer simulations agree with each other would be enough to ensure safety. They ran the same scenarios on two different simulators and compared their results. The two computers never disagreed. They both passed the same tests and both missed the same failures. This revealed a shared blind spot: if the simulation is built on a certain set of rules, both simulators will follow those rules and miss the same real-world problems. Only the physical measurement, the camera watching the real light, exposed the error. This finding suggests that running multiple simulations is not a substitute for testing in the real world. If the simulations are built on the same assumptions, they will all be wrong in the same way.

Finally, the researchers looked at how the design of the agent itself affected these results. They found that changing the way the agent was configured—specifically, how it connected to the software that ran it—could fix the problem entirely. In one case, switching the configuration made the agent's actions match the simulation's expectations with an AUC of 1. This showed that the problem was not always with the agent's intelligence, but with the invisible connections between the agent and the house. The "auditability" of the agent, or how well we can trust its results, depended on these technical details.

The work of SimVerity does not offer a magic fix that makes all simulations perfect. Instead, it offers a new way to think about safety. It suggests that we should stop treating a simulation's green light as a final guarantee. Instead, we should treat it as a starting point that needs to be checked against the messy, slow, and unpredictable reality of a physical home. The researchers propose a simple decision process: clear the agent only if the evidence supports it, hold back if the evidence is missing, or escalate the issue if the simulation says it is safe but the real world says it is not. In a world where our homes are becoming smarter, the most important step is not just building better agents, but building better ways to know when they are actually ready to live with us. The green check from the computer is a promise, but as this study shows, that promise must be verified by the slow, steady truth of the physical world.