Local use of open protein-structure models is attractive for
researchers without CUDA cluster access, but hardware-viability claims
need target-level evidence rather than anecdotes. We report an 8-target
reproducibility benchmark of SimpleFold-3B on a 128 GB M3 Max
workstation, using a direct local CLI route and public audit artifacts.
Phase 1A froze the target set (13CZ:A, 13EE:A,
13FT:A, 11BC:C, 12OY:A,
9H8N:A, 9K9X:A, 9K9Y:A) before
the four prospective targets were executed. All 8 targets produced
predictions and scores. Across the frozen set, median wall-clock time
was 395.96 seconds, median approximate peak RSS was 34.08 GB, median
sequence-aligned lDDT was 0.8597, and median TM-score was 0.9218. The
three long-bin targets (565-604 aa) completed without OOM at 34.02-34.64
GB peak RSS. Selective repeat controls now cover the 13CZ:A
route/cache anchor plus the reviewer-sensitive 11BC:C
hard/coverage target and 9K9Y:A long-bin target. This is
not a model leaderboard, not a reproduction of the SimpleFold CASP14
headline benchmark, not an official AlphaFold 3 evaluation, and not a
biological generality claim. Under this exact model, hardware, execution
route, target set, and scorer, local SimpleFold-3B inference was
operationally viable and auditable; the public source snapshot and
Zenodo archive are available at
https://github.com/Theovia/eigen-reprod-atlas-phase1a/releases/tag/v0.1.2-phase1a
and https://doi.org/10.5281/zenodo.20275055.
Keywords: protein structure prediction, SimpleFold, Apple Silicon, MLX, reproducibility, benchmark provenance, Cloudflare Workers, single-researcher.
This abstract intentionally does not claim SimpleFold paper reproduction or multi-model benchmark results. Those claims require a different no-fixture published-metric comparison and the later Phase 1B expansion.