robotruth is an open source Python library and CLI that tells a robot lab whether a policy change is real. Contract checks, honest statistics, drift fingerprints, a calibrated outcome judge and a runtime guard.

Which Vercel product did you use? What new capability did it give your team, and how did that shape the product you launched?

Vercel based hosting is used in robotruth

Robot CI for learned robot policies. It tells a robotics lab whether a policy change is real, before the robot, the eval or the launch demo tells them the hard way. Six modules: a fail-closed contract checker that proves the policy you deployed is the one you evaluated, honest statistics that refuse to print a rate without its interval, episode records and fleet metrics, drift fingerprints for the cell and the arm, a calibrated outcome judge that abstains instead of guessing, and a runtime guard. Everything in it is measured on public data and published with the scripts that produced it, including the runs that contradicted earlier claims. The headline finding is a correction to my own tool: ten failure detectors, nine of them from the recent literature, all sit at chance on the task failures real robots actually had, while catching injected execution faults almost perfectly.

Which Vercel product did you use? What new capability did it give your team, and how did that shape the product you launched?

Vercel based hosting is used in robotruth

robotruth

It tells whether a policy change is real?

It tells whether a policy change is real?