Preliminary controlled results, August 2026

First public release of the benchmark apparatus and preliminary evidence.

Accepted cohorts:

• input-keypress-architecture-v3: N=1 descriptive efficiency win at equal correctness
• package-cache-is-empty-v2: N=1 fixed-overhead loss at equal correctness
• settings-label-roundtrip-v4: N=2 null correctness result with mixed, mostly negative efficiency

The release also retains seven excluded run versions with machine-readable reasons, 43 redacted ATIF trajectories, attempt receipts, verifier evidence, logs, frozen fixtures, preregistrations, and reproduction instructions. The cohorts are not pooled and do not support a general productivity claim.