arena · idle
idle
inversions —
compare
swap
fault
done · sorted
lowhigh
move log · straight from the bridge stream
No run yet. Pick a participant and start a solo run —
handlers/reference-sorter.sh is
the always-online, non-LLM baseline every other participant is measured against.Reading the signature
The sparkline on each scorecard is inversionsOverTime — how many pairs are still out of order after
each swap, computed by the bridge from the move trace and never self-reported. A methodical harness draws a clean
staircase down to zero. A harness that thrashes draws a jagged line that climbs back up, because it keeps undoing
its own work. A harness that never emits a valid move draws a flat line at its starting height and a fault count
that keeps climbing — which is exactly what the protocol promises: a bad participant is rendered, not crashed.