The experiment, with its evidence.
An experiment engine
for physical AI.
A robot task fails after a deployment changes. Nisayon records experiments that test candidate corrections, retains the regression, and returns unresolved when the evidence cannot support a decision.
Read the experiment
The instrument presents recorded Panda Lift experiments. Select a case, compare reference and regression runs, and move through the retained timeline. Actions, observations, timing, and task progress remain tied to the selected row.
The 3D scene is a reconstruction from retained configuration. Original readings remain separate. Playback shows recorded data; it does not execute a robot or report a live experiment.
Inspect the decision
Open Decision for the diagnostic result, evaluator findings, trial record, and separate confirmation record. A completed process does not establish a successful task or an accepted correction.
A candidate must be frozen and checked on fresh conditions before acceptance. Missing confirmation, unknown policy state, invalid measurements, and observed failure remain explicit.
Follow the source
Open Engine for the task contract, diagnosis, measured costs, exact evidence files, and implemented interfaces. Generation manifests bind files by SHA-256 and byte length. Integrity checks establish file agreement; they do not prove physical truth.
The public site reads exported records. Execution and workspace tools run locally. Generalization to unseen faults, physical safety, and a cost advantage are not established by these development cases.