ProgramBench — Fa results (⭐ self-reported)

Fa's results on facebookresearch/ProgramBench — rebuild a complete codebase from a compiled binary + its docs. Pass@1: one attempt per instance — the Solved column (solved/attempted) is the pass@1 fraction the official board ranks by; Test-pass % is the mean share of scored tests that passed per instance. Measured by our own bench-programbench runs. Every row is self-reported and never auto-submitted to the official registry; smoke rows are pipeline proofs, not leaderboard rows. Each row links its run and the analysis artifacts (trajectories, eval logs).

No ProgramBench results recorded yet — the ledger fills on the first green bench-programbench run.

_{Source: bench/leaderboard/pbench_results.json, rendered by bench/leaderboard/pbench_render.py (gh-182) — deploy-time product of fa1.dev.}