Glossary

The competition

• Unit — one task and two scenes of it: the demonstration's and the one the policy is scored in. Both sides of a duel play it.
• Run — one of a task's units in a duel; each task gets 10.
• Duel — a challenger and the king on an identical unit list.
• Genesis — an entry taking the empty throne, scored on its own units: the first entry, and the first after a king is vacated.
• King — the model holding the crown.
• Champion — one of the four most recent kings, each paid a share of Vector's emissions.
• Score — a model's mean success rate over the tasks with at least one scored unit.
• Margin — how far the challenger's score must clear the king's: +3.0 pts.
• Void unit — a unit the harness could not run, scored for neither side.
• Void duel — a duel with more than 20% of its units void: it is not published and moves nothing, and the challenger's submission is spent.
• Refused — an entry whose repository, digest or weights fail the checks; it never duels.
• Duplicate — an entry whose weights are byte-identical to an entry already admitted.
• Missing — a repository the Hub does not show, or a Hub that does not answer, when it is looked up: a challenger is dropped from the queue, a king is vacated.
• Vacated — a king dethroned without a duel because its repository was gone; its place among the champions stays, and its share is burned.
• Forfeit — a king whose repository disappears during a duel; every king unit counts as a failure.
• Seed block — the finalized chain block whose hash dealt a duel's units, read when the duel starts and after the challenger's commitment.
• Commitment — what a hotkey writes on chain to enter: the repository, the revision and the sha256 of its weights. One per hotkey.
• Contract — spec.json, the file every protocol number is read from; a duel's id includes its version.

The model

• Demonstration — the expert doing the task in another scene: three camera streams and the arms' poses at every step.
• Endpose — a 16-number row: per arm, end-effector position, orientation quaternion and gripper.
• Prompt — the demonstration as the network holds it: one token per 15 steps, plus sink tokens.
• Sink tokens — four learned tokens in front of the prompt that attention can always rest on.
• History — the frames a prediction reads: the current one and the one before.
• Action chunk — the 16 actions one prediction returns; the first 12 are executed.
• Rot6d — a rotation as the first two rows of its matrix, the form the network predicts.
• Adaptation gain — how much a model's score drops when its demonstrations are swapped for ones of other tasks: a measure of whether it reads its context at all.