ScaleArena

Leaderboard ?

Predicted vs measured ?

Kill-criterion details, pair by pair

Pre-registration note (what was declared before the battery ran)

1 · Write the algorithm

A new row is a file. It implements decide(Observation) -> Action(delta); the harness clamps the delta to the card's step cap, n_max and boot concurrency.

Row name

Lower-case letters, digits and dashes; becomes plugins/<row>.py and the arena row name.

Label

One line shown beside the row everywhere in the arena. It is written into ENTRANT["label"] in the source below as you type.

kwargs (JSON)

Constructor keywords, written into ENTRANT["kwargs"]; the card adds delta_s and n_max.

Source · plugins/<row>.py

Pre-filled from the plugin template. Frozen rows (keda, ratefwd, ratefwd-hw) cannot be overwritten.

overwrite an existing file

Rows added to this arena

Plugins under plugins/. Remove takes the row off every leaderboard of the selected card (its run directories and journal lines go, the card is rescored) and deletes the file unless another card still has runs for it. Frozen rows are cited by pre-registered batteries and stay.

3 · Pre-register for the cluster

Writes tier2/specs/<battery>.spec.json with the predictions as its arms_note?, then prints the commands. The UI never touches the cluster.

Battery id

Hardware card

Families

Seeds / windows

Duration per run (s)

Extra arms

Commands to run on the head (operator)

2 · Result

Nothing validated yet. Write the row on the left and press Validate: the file is written, imported in a subprocess and asked to decide() on synthetic observations.

Simulator job

Simulator run Card Trace Row

Cluster run Battery Run