# KaiB demo data — provenance

This directory contains a concatenated demo asset for the SF-Cluster Colab
notebook. It is derived from the SF-Cluster Phase II benchmark's KaiB
`diverse_sf` arm and the FrustrAI-Seq per-residue Frustration Index (FI)
outputs.

## Files

| File                  | Shape / size           | Description |
|-----------------------|------------------------|-------------|
| `KaiB_filtered.a3m`   | 364 records, L=91      | Subset of the KaiB filtered MSA. Query (`>101`, UniProt Q79V61 residues 5–95) is row 0. Lowercase insertion-state letters preserved. |
| `KaiB_fi_matrix.npy`  | (364, 91) float32      | Per-residue FI matrix. Row `i` corresponds to record `i` in the A3M. |
| `KaiB_seq_ids.txt`    | 364 lines              | One short sequence ID per line, in the same order as the A3M / FI matrix. |

## Provenance (internal source paths omitted)

Derived from the SF-Cluster Phase II KaiB `diverse_sf` benchmark arm: the
filtered KaiB MSA (depth 6821, L=91) and the per-subset FrustrAI-Seq FI
artifacts (12 subsets, each a `(32, 91)` float32 `fi_matrix.npy` with
`metadata.json`, `fi_residual_matrix.npy`, `entropy_matrix.npy`) mapped back to
the 12 subset A3Ms to recover sequence IDs. Internal dev-repo paths are omitted
from this public demo asset.

## Construction recipe

1. For each of the 12 `diverse_sf` subsets, load `fi_matrix.npy` ((32, 91)
   float32) and the corresponding `diversesf_KaiB_KaiB_seed{NNN}.a3m`.
2. Concatenate rows in subset-index order; track the parallel sequence-ID
   list from the A3M records.
3. **Deduplication policy**: first occurrence wins. A sequence ID seen in an
   earlier subset is skipped (both in the FI matrix and the ID list). This
   reduces 12 × 32 = 384 raw rows to 364 unique rows.
4. Extract the corresponding sequences (with their full headers and
   lowercase insertion states) from the filtered MSA, preserving the order
   established in step 3. The query (`>101`) is always row 0.

All 364 unique IDs were found in the filtered MSA (0 missing).

## Models

- **FrustrAI-Seq weights**: HF repo `leuschj/FrustrAI-Seq`, commit
  `ee5a01a29fde00630f4a1157f0e6cb8343ac434b`. Inference in fp16 with LoRA
  adapters merged.

## License

This demo asset is released under MIT alongside the SF-Cluster OSS package.
The KaiB sequence (UniProt Q79V61, *Thermosynechococcus elongatus*) and its
MSA neighbors are public-domain sequence records via UniRef100 / Mgnify;
no proprietary structures are included. FrustrAI-Seq outputs are derived
features (floating-point FI values) and are released by the FrustrAI-Seq
authors under their own license — see
https://huggingface.co/leuschj/FrustrAI-Seq.