Continual-ARCby unify

ReplayARC-AGI-1 · Fresh · GPT-6 Luna, low

OpenClaw, as shipped

Every learning mechanism the harness ships switched on.

Held-out tasks, budget used
56.1 solved 58%
Model spend
$3.87 1,905 calls
Instances
100 14.7 hours

Budget used is the share of the cap of 16 that an instance was charged, averaged over the instances: a solved instance costs 1 for each wrong answer and each demonstration, an unsolved one all 16, so lower is better.

Loading the run