# Embedding spaces — provenance and reproduction

The benchmark ships four precomputed embedding caches. **All of them were
computed by us from public model weights applied to the public 490-sequence
cohort** (`cache/k562_experiments/cohort/cohort_doublet_filtered.csv`,
UniProt reviewed human sequences; headers include `sp|ACCESSION|NAME_HUMAN`).
No embedding file is taken from a third party — what is third-party is the
**model weights** (public checkpoints, exact paths below).

## Summary table

| File (cache/embeddings/) | Shape | Model checkpoint (public) | Pooling / normalization | Computed by |
|---|---|---|---|---|
| `k562_esm2_650m_491x1280.npy` | (490, 1280) | `facebook/esm2_t33_650M_UR50D` (HF) | residue mean, L2-normalized; sequences >1022 aa split into 1022-aa chunks, length-weighted mean | us |
| `esm2_3b_490.npy` | (490, 2560) | `facebook/esm2_t36_3B_UR50D` (HF) | same protocol as above | us |
| `protrek_490.npy` | (490, 1024) | `westlake-repl/ProTrek_650M` (HF) + code from `github.com/westlake-repl/ProTrek` | `get_protein_repr` (official API; the model returns L2-normalized 1024-d protein representations) | us |
| `rpe1_protrek_patched_1091x1024.npy` | (1091, 1024) | same ProTrek checkpoint | same API; base 1087-TF cache + 4 missing TFs patched with a gated sanity check (below) | us |

Every file's SHA-256 is pinned in `data/embedding_checksums.txt` and verified
by the test suite.

## Reproduction

```bash
# weights (one-time, ~3.6 GB for ProTrek; ESM2 auto-downloads from HF)
bash data/download_models.sh models

# rebuild any cache and check it against the pinned SHA-256
python scripts/recompute_embeddings.py --model esm2_650m
python scripts/recompute_embeddings.py --model esm2_3b
python scripts/recompute_embeddings.py --model protrek --protrek-src models/protrek_src
```

The script exits non-zero if the recomputed checksum does not match, so the
caches are verifiable rather than trusted.

## The RPE1 patched cache — what "patched" means

`rpe1_protrek_patched_1091x1024.npy` serves the RPE1 raw-data pipeline. Its
base is a 1087-TF ProTrek cache computed with the identical checkpoint and
API. Four training TFs were missing from the original cache and were added
in 2026-09 with a **gated sanity check**: 6 TFs already present were
re-embedded first, and the patch was written only because all cosines vs
the existing cache were ≥ 0.9999 (measured 1.0000). Full receipt with the
per-TF alias resolution (UniProt accessions, readthrough-gene exclusion):
`data/splits/embedding_patch_receipt.json`.

## Using a different encoder

The benchmark is encoder-agnostic: compute any (490, d) embedding for the
cohort sequences with your own model, L2-normalize rows, and pass it via
`load_benchmark(embedding=...)` (or your own array — `SplitData.phi_*` are
plain numpy). ESM2-650M is the default because it is the cheapest
reproducible reference point; conclusions in the associated paper were
verified to be flat across all three encoders above.