Accuracy

Before every release, JapanFold's output is checked for parity against each model's official reference implementation. The method, the reference backend used for each leg, and the per-model evidence live in tt-bio's implementation-parity docs.

A leg compares the device against the reference on the same input, and passes when the device sits inside the reference's own seed-to-seed noise floor. Where a metric sits outside that floor, the parity docs carry the measured number and the root cause: these are precision and sampling-noise floors, not defects in the port.

This measures the port, not the science. It compares the device against the reference implementation on the same input, not against experimental structures, and it inherits whatever the published checkpoint is worth. OpenFold3's weights are a preview checkpoint trained well short of the full AlphaFold3 schedule, so read its confidence scores before trusting a prediction.

The 2026-09-24 engine update

The same request with the same seed now returns a slightly different structure. We folded DHFR (187 residues) on each model with the old and the new engine, seeds 0 and 1, and measured the CA RMSD after superposition. Every model moved less between engines than it moves between two seeds on one engine:

Model	Old vs new engine	Seed 0 vs seed 1
Boltz-2	0.11 Å	0.30 Å
ESMFold-2	0.10 Å	0.12 Å
ESMFold-2 Fast	0.06 Å	0.72 Å
OpenBind-0	0.07 Å	0.31 Å
OpenDDE	0.05 Å	0.21 Å
OpenDDE AbAg	0.04 Å	0.18 Å
OpenFold3	0.04 Å	0.28 Å
Protenix-v2	0.10 Å	0.97 Å
RoseTTAFold3	0.03 Å	0.07 Å

ESMC 6B embeddings are unchanged except at lengths whose token count is a multiple of 32, where the old engine was wrong (see the changelog).