raftbioworks/boltz2_accuracy_architecture_20260919_032813

Boltz2 accuracy and architecture study

Correction: The original Cα scoring selected the wrong atom. Use this revision’s corrected measurements and see the correction notice before using any prediction archive. Raw archived ca arrays are superseded; recover Cα from full coordinates and token_to_center_atom.

Measured small-protein ablations, an evaluation RNG/precision repair for the local Boltz quotient adapter, and a cyclic Fourier triangle-contraction prototype. No new weights were trained. This is not evidence of accurate 100k folding.

See the report, 38 architecture ideas, and measurements. References are public PDB structures. checkpoint.json identifies the unchanged checkpoint already publicly archived in our preceding audit.

results.tar.gz contains per-case/seed coordinates, input features, status, RNG hashes where recorded, and logs. Failed harness attempts are retained. The old shared_rng field denotes model-entry RNG only; use diffusion_rng_sha256 for the corrected comparison. patch-proof-final uses only the evaluation patch; patch-proof-reference uses both patches. Do not confuse their legacy mode label with the unpatched source.

For reproduction, install the Anthropic Boltz2 kit at f4f62fa6592ae4938d49b1757bea0cfeff9f468e in its separate Python 3.11 / PyTorch 2.12.0+cu130 environment. The launcher references an existing Modal image/volumes in our account; other accounts should build the kit image and mount their own unchanged Boltz2 cache. Set /study to the reproduction scripts and place the chosen original_source/boltz_zeus or fixed_source/boltz_zeus first on PYTHONPATH. The final patched source already includes both patches. point.py uses W&B and a Modal volume commit; adapt those account-specific destinations. Biological runs use 3 recycles, 200 sampling steps, empty MSAs, one sample and seeds 42/43.

All GPU timing measurements are synchronized forward timings rather than download/startup/end-to-end service latency. Native comparisons use complete-assembly alignment, coordinate coverage and pair-weighted C-alpha lDDT, not DockQ. Two seeds on a small selected panel are not an equivalence study. Source code is supplied under its existing licenses; scientific facts and PDB-derived coordinates retain their source provenance.