ProgramBench proof

Four difficult programs. Four retained completions.

ProgramBench asks an agent to rebuild a whole program from an executable and documentation—without source access, internet, or decompilation.

4retained tasks
completed

For context, the official 200-task leaderboard reported a best fully resolved rate of 0.5% on July 24, 2026: one task.

retained taskinstancepassing tests

Hyperfinesharkdp__hyperfine.327d5f4298 / 298

Code Minimapwfxr__code-minimap.0ddeea5369 / 369

Brotligoogle__brotli.b3dc9cc606 / 606

Nominoyaa110__nomino.f892499338 / 338

Inspect, don’t trust

The rule, classifications, custody hashes, and reproduction surface are public.