ESMFold2: result.plddt is on [0, 1] but result.complex.to_mmcif() writes B-factor on [0, 100] — intentional?

[korbonits]

Wanted to flag what looks like a scale inconsistency between two surfaces of the same ESMFold2InputBuilder().fold() return object, and check whether it's intentional or worth a docs note.

Tested against commit: 81b3646c9429ea8458918415ad6a46178cb59833 (the pin in modal-labs/modal-examples for protein-folding/esmfold2.py). I haven't tested against main or other revisions, so the behavior may differ there.

Setup: default biohub/ESMFold2 weights, single 53-residue chain (MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQAPILSRVGDGTQDNLSGAEK), num_loops=3, num_sampling_steps=50, num_diffusion_samples=1, seed=0, on a Modal H100.

Observation:

result = ESMFold2InputBuilder().fold(model, spi, num_loops=3, num_sampling_steps=50,
                                     num_diffusion_samples=1, seed=0)

result.plddt.min().item(), result.plddt.max().item()
# (0.4218, 0.6786)   ← fractional, on [0, 1]

cif = result.complex.to_mmcif()
# B-factor column min/max: 42.18 / 67.86   ← on [0, 100]

Same metric, same call, two different scales depending on whether you read the raw tensor or the mmCIF text. The B-factor scale matches the AlphaFold / ESMFold-v1 convention (so pymol spectrum b, 50, 90 "just works" on the written file), but the tensor doesn't — and result.plddt's docstring doesn't mention the scale, so a downstream caller plotting from the tensor sees values clustered around 0.5 where they'd expect 50.

Question: is this intentional (raw tensor as "model output," to_mmcif() as "presentation layer" that conforms to PDB B-factor convention), or unintentional? Either way, a one-liner on result.plddt clarifying the scale would save the next integrator the round trip — happy to send a docs PR if useful.

Receipts: integrated this into Sheaf v0.11 (a serving layer for non-text foundation models); the full verification trail with the H100 smoke output is in ADR-0001 if it helps reproduce.