How it works

Detection against the machine's own normal

TrialONE does not look faults up in a library. It measures a recording against a reference: the rest of the same recording, earlier recordings of the same machine, or a healthy recording you name with --healthy. The same math runs on a file and on a live stream, and it finds faults nobody has named before.

The strongest detector compares repeated operations with each other. That is why TrialONE is built for machines that repeat their work: mobile robots, manipulators, quadrupeds and humanoids. The detection is reproducible: the same recording gives the same findings.

Explanation: what the signal is, and why it moved

Channel names are whatever your team called them, so meaning is not guessed from a name table. An agent layer reads the numbers, the log lines and, with --project, your source tree, and proposes a physical story. The story has to commit to predictions that must also hold if it is right, and the engine checks each one against the data. A story whose predictions fail is reported as refuted, with the failures named.

The explanation never changes what was detected. With --no-explain you get the deterministic half on its own.

Three zones per finding

An explanation can raise a zone, never lower it. Every report says how many findings stayed undecided, out of how many.

Routing: software or hardware

Every finding goes to the person who can fix it. A software fault goes to a coding agent, with files and lines. A physical fault (a connector, a part, the mechanics) goes to a technician at the machine. In a robot learning dataset, a fault that the policy's own commands carry goes to the person who trains the policy.

Proof: the next recording decides

A person or a coding agent claims a repair; the robot proves it. trialone verify measures a new recording against the earlier diagnosis and rules per finding: gone, still there, or new. A lower peak is not a pass. See Proving a repair.

What it does not do yet

• A fault on one single channel, with nothing else moving at the same time, produces no finding on the live path. Faults across several channels are found within a second.
• Explanations on the live path are leads, not verified answers. Predictions are checked against recordings, not against a running stream.
• A fault present in every operation of a recording is invisible without a healthy reference. Pass --healthy.
• Camera frames and point clouds are not analysed. Numeric channels, log lines and crash reports are.
• Drones are read, but every flight is different, so there is no repeated operation to compare against. They are not a focus.