Continual-ARC

ReplayARC-AGI-1 · Workstreams · GPT-6 Luna, high

Prime Agent, told to keep a program library

Told to keep one tested program per task and run it first when the task returns.

Stream, budget used
9.8 solved 99%
Epilogue, budget used
20.3 solved 96%
Model spend
$2.37 2,460 calls
Instances
225 5.8 hours

Budget used is the share of the cap of 16 that an instance used, averaged over the instances: a wrong answer and a demonstration each cost 1, so lower is better.

Loading the run