Computational benchmark demonstrates 24.2% accuracy on ARC-AGI-2 under a zero-dollar budget, suggesting failure memory and candidate diversity outweigh raw model scale.
We document a zero-marginal-cost approach to ARC-AGI-2 that reaches 24.2% on the public evaluation set (29 of 120 tasks solved at >=0.95 pixel accuracy: 15 exact, 14 in the 0.95 - 0.999 frontier band) - 100.7% of the current open SOTA (NVARC, 24%) - using handcrafted grid primitives, one-shot LLM program synthesis with execution verification, a persistent near-miss library, and hint-injected repair. The load-bearing claim: under a hard $0 budget, the binding constraint is not model intelligence but candidate diversity and failure memory. We measured 37 prototype iterations; every number in this writeup regenerates from receipts on disk. The companion notebook reproduces Prototypes 07/09 and the library pipeline.
No takes yet. Share an insight, caveat, or question.
R. E. W. Kho (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: