A correct-looking agent plan is not a verified physical result

An eight-channel laboratory pipette can be commanded to aspirate from a trough it cannot physically enter. The instruction may be syntactically valid, and the agent may even be able to explain it. Neither fact allows the robot to perform it. Google Research's October laboratory disclosures and a separate simulated-driving study show why checking an agent's plan, its permitted action and the observed result are three different jobs.

VERID maintains a stateful digital representation of a workcell: deck geometry, robotic arms, pipette tips, vessel volumes and reagent identities. The reported eight-channel aspiration was rejected because the selected trough was too narrow. The agent replanned before the invalid command reached physical equipment. The published examples also include recovery from a plate-reader failure and experiments to diagnose fluid-handling biases. They support a particular property: known geometrical and state constraints can be checked outside the language model.

A digital twin does not make unmeasured physical effects disappear. A residual liquid film, the accuracy of a camera or an incomplete model of a reagent can still undermine a technically executable protocol. This is a limit of the modeled physical state, not proof that simulation is worthless.

Google's aptamer design experiment failed at a later stage. In its second design round, the agent enriched G-quadruplex motifs that increased raw binding signals but also produced adhesion to beads and other targets. The reported myoglobin specificity dropped from 75.0% to 19.1%, and bead-corrected hit rate from 10.2% to 3.7%. Once the measured artifact was recognized, the next round recovered specificity to 72.4% and raised the on-target hit rate to 33.5%.

The pipette example concerns whether a command can be executed. The aptamer example concerns whether an executed experiment measures the intended objective. Both can produce an apparently successful software trace while answering different scientific questions. The real assay result, rather than the agent's internal confidence, was the evidence that prompted correction.

SafeDriveVLA evaluates another boundary: the user or navigator can request a maneuver that conflicts with driving constraints. In the paper's selected simulated scenarios, a world-model-guided alternative reduced collision counts from 60 to 38 relative to the tested safety-oriented baseline. The denominator and scenario design matter. These are simulated collisions in a benchmark, not an estimate of road accident rates or a certification for real vehicles.

An operator should be able to reconstruct the requested goal, the constraints actually enforced, the actions that reached external systems and the independent measurement of their outcome. The evidence may include a simulator rejection, equipment logs, a binding assay or a driving trace, but the fields are not interchangeable. An appealing final response cannot establish that an action was authorized, that hardware executed it or that the resulting measurement represented what the human wanted.

The shared design lesson is narrower and more useful than a blanket promise of safe autonomy: models can search for alternate plans, while external state checks and independent outcomes remain responsible for deciding which plans are executable and which results count.