Feature request: expose reproducibility metadata and full interface confidence for structure NIMs

Hi BioNeMo team,

For reproducible protein-complex research, could the OpenFold3 and Boltz-2 NIM skills/API documentation make the effective model configuration and complete confidence outputs more transparent? The current skills document useful request controls (for example, diffusion samples and Boltz-2 recycling/sampling steps), but I could not find a documented explicit random seed, immutable model/checkpoint identifier, or a guaranteed full PAE/inter-chain PAE and chain-pair interface-confidence response contract. If these already exist, pointers to the exact fields would be very helpful.

A useful research-oriented request/response contract would include:

• An explicit seed or documented sampling behavior, with seed/sample IDs returned for each structure.
• The effective parameters and defaults actually used, not just the submitted fields.
• A model/checkpoint revision or digest and container/image digest, with hosted/local version parity documented.
• Per-residue pLDDT, full PAE (especially inter-chain PAE), pTM, ipTM/chain-pair interface scores where supported, and a clear description of any score that is not available for a given model.
• A machine-readable run manifest/example that preserves request, response, MSA/template provenance, and all samples.

Would you consider documenting these fields or exposing them where the NIM API does not yet provide them? That would make self-hosted NIMs much easier to use in auditable scientific studies, where interface scores must be compared across seeds, controls, and construct boundaries. This is a general feature request; no private sequences or experimental data are included.

Start by locating the OpenFold3 and Boltz-2 NIM skills/API documentation and compare the documented request controls with the requested reproducibility and confidence fields. Done means the documentation or API clearly defines effective parameters, seed and model/container identity, confidence outputs, unavailable scores, and a machine-readable run manifest with provenance.