README: contents at a glance, provenance, and who consumes this pool

Records what the deposit actually holds (15,018 metadata rows, 10,884 predicted +
369 native PDBs, 249 epitopes, 81 alleles) and which columns are measured versus
computed -- the pairings are VDJdb records, the coordinates and every confidence
metric are model output, and the derived angles track AlphaFold confidence rather
than being independent of it. Also names the downstream consumer: the TCRen2
receptor benchmark draws its real-binder positives from here.

README.md

@@ -29,4 +29,38 @@ Metadata for all structures is concatenated with corresponding VDJdb record meta
 * `ranking_confidence`, `plddt`, `ptm`, `iptm`, `tcr_pmhc_iptm` - TCRmodel2 quality metrics (for more information please visit [TCRmodel2 webpage](https://tcrmodel.ibbr.umd.edu/help)).
 * `scanning_angle` and `pitch_angle` - scanning and pitch TCR:pMHC structure angles calculated with STCRpy.
 * `is_native` - whether the structure is experimentally acquired (from PDB) or predicted.
-* `tcr_pmhc_hash` - unique TCR:pMHC record hash for matching predicted structure PDBs.
+* `tcr_pmhc_hash` - unique TCR:pMHC record hash for matching predicted structure PDBs.
+
+## Contents at a glance
+
+| | |
+|---|---|
+| metadata rows | 15,018 (11,234 distinct `tcr_pmhc_hash`) |
+| predicted / native rows | 14,612 / 406 |
+| PDB files shipped | 10,884 predicted (`pdb_files.tgz`) + 369 native (`pdb_files_native.tgz`) |
+| epitopes / MHC alleles | 249 / 81 |
+| MHC class | 14,915 MHC-I, 103 MHC-II |
+| species | 14,928 human, 90 mouse |
+
+A metadata row without a matching PDB is normal — the file set is the subset that modelled and passed
+QC. Join on `tcr_pmhc_hash`; predicted filenames are `aligned_aligned_<hash>.pdb`, natives are
+`aligned_<pdbid>.pdb`.
+
+## Provenance
+
+- **TCR:epitope pairings are EXPERIMENTAL** — VDJdb records, carried through with the full VDJdb
+  "full table" columns (`reference.id`, `method.*`, `meta.*`, `vdjdb.score`).
+- **Predicted structures are COMPUTED** (TCRmodel2/AlphaFold). `ranking_confidence`, `plddt`, `ptm`,
+  `iptm`, `tcr_pmhc_iptm` are the generator's own confidence metrics, not measurements.
+- **Native structures are EXPERIMENTAL** (RCSB), re-processed into the same canonical frame.
+- **Derived columns** (`num_contacts`, `scanning_angle`, `pitch_angle`) are computed from the
+  deposited coordinates — for predicted rows they inherit the model's error, and the angles in
+  particular track AlphaFold confidence rather than being independent of it.
+
+## Used by
+
+The TCRen2 receptor-ranking benchmark (`isalgo/tcren_structures`) draws its **real-binder positives**
+from this pool — filtered to `is_native == False`, restricted to hashes present in `pdb_files.tgz`,
+and with TCRvdb pairs excluded — then matches them per epitope against the decoys in
+`isalgo/tcren_structures::vdjdb_negatives/` and `::immrep23_negatives/`. See
+`isalgo/tcren_structures::vdjdb_binder_benchmark/README.md`.