Collect structural observations

Read AUTHORITIES.md, extract and proofread. The structural aggregate is chunks/PDB_Database.tsv; retain the publication reference on each row and the accession in meta.structure.id. A multi-PDB paper has one intake issue, not a chunk for each structure.

Establish the source set

Read the paper, supplements and PDB primary citations. Separate newly deposited structures, reused templates, new assays and predicted models. Inventory every accession with its reported receptor, peptide, restriction and assay outcome. A cited template is not a new experimental observation. Distinct deposited structures can remain distinct observations under one publication.

Fetch mmCIF coordinates and entry/polymer metadata. Check each response independently; a failed first download must not hide the other accession results. Unavailable coordinates are a source blocker: keep the issue open with the accessions and admission condition. Do not make empty or sequence-free chunks. Preserve downloadable source attachments when shortening issue text.

Use tcren locally

Use ~/vcs/code/tcren and read its AGENTS.md and ~/vcs/code/tcren/skills/tcren/SKILL.md. Install it into an isolated extraction environment when needed. This is a curation tool, not a source-checkout replacement for this repository's published build dependencies.

Parse the complete set with tcren.structure.parse_structure, annotate with tcren.annotation.batch.annotate_batch, classify with the resulting precomputed records, and run tcren.mhc.annotate_mhc_batch once over the set. Annotation is batched; never put mmseqs calls in a per-structure loop or a process pool. Resolve reference roots explicitly and use a separate output/data root for the extraction.

Keep chain identifiers, literal polymer sequences, coordinate sequences, annotations and allele candidates in the extraction proof. Inspect chain roles and the peptide in the groove; short protein fragments are not automatically epitopes. Single-chain pMHC constructs need their reported peptide separated from linkers, beta-2-microglobulin and the MHC chain.

Check what the coordinates omit

Compare coordinate residues with the deposited complete polymer sequence. Unresolved residues can shorten a junction without changing the submitted construct. Recover missing residues only from the complete deposited sequence or an explicit source sequence; do not copy a template's CDR3 into an engineered receptor.

VDJdb uses junction_aa, including Cys104 and Phe/Trp118. tcren CDR3 region markup excludes the anchors. Check the FR3/FR4 sequence and the J motif; adding C and F by concatenation is not a valid conversion. Keep inferred gene/allele calls distinguishable from published assignments.

An MHC groove sequence can match several alleles equally. A top hit, including a null-expression allele, is not proof that the experiment used it. Use the reported construct and primary PDB annotation to resolve the allele; retain reported resolution or flag a material contradiction. Inspect engineered MHC substitutions and nonstandard peptide residues explicitly.

Check specificity and publish

A covalent clamp, forced docking model or common orientation does not establish specificity. Review uncoupled binding and functional assays. Failure to activate does not erase separately measured binding; preserve the distinct assay outcomes. Engineered receptors require their own literal sequences and evidence, not the parental receptor's assignments. Culture of a reconstructed receptor does not establish culture before sequencing.

Format supported observations in the current chunk schema. Keep unavailable, unassigned, explicitly negative and chemically modified observations in their appropriate source dispositions. Use duplicates to compare complete observations and publish for a PDB/data-issue branch, registry reconciliation, separate validation and release-aware checks. Close the intake only when its material scope is accounted for; a PDB citation match alone is not complete curation.

Reported peptides and predicted cores

Preserve the complete assay peptide and source-supported MHC/provenance. Use standalone vdjdb assess-epitopes during review when predicted presenters, binding registers or TCR-facing representations help resolve a comparison. Follow the assessment guidance. Predictions do not supply a missing experimental epitope or establish equivalent recognition. A predicted core can omit required flanks or insertions and does not replace a structure's full ligand.

Audit hosted coordinate files against current records

Use the independent audit after current tables have been assembled:

uv run python processing/proofread_structure_files.py --structures out/inputs/structures \
  --tables out/tables --out out/reports/structures --fetch

proofreading/structure_inputs.tsv pins the hosted raw files by revision, size and SHA256. Check its current download sizes and available disk first. --fetch stages those inputs separately; omit it when checking an existing local mirror. The audit verifies every digest and streams all declared archives without extraction. It checks current links against reviewed coordinate-sequence metadata and verifies coverage for every native accession in current records. CI does not independently re-extract chain sequences. The current scope is 367 native accessions across all chunks; the older 360-accession count covers only the PDB aggregate chunk. No chunk or coordinate is edited.

Inspect structure_files.parquet, native_structure_files.parquet, supplemental_structure_files.parquet and structure_files.json in the report directory. Keep original/repaired sequence matches, exact receptor:pMHC matches and historical lineage links distinct. A historical association is not sequence proof. The supplemental catalogue contains 11,911 distinct coordinate contents, with 11,388 new and 523 existing contents. Its 16,118 sequence-state links connect 4,777 structures to 7,292 current records; they do not count independent experiments. Use its README and verification report for dispositions and input identity. Audit supplemental raw assets and their current links separately. Ignore input positive/negative, pairing and validation flags for linkage. These inputs identify receptor/pMHC sequences and files, not binding outcomes. Outcome labels come only from curated negative chunks or reviewed 10x removals where applicable. Require literal coordinate/sequence/context agreement regardless of flags; keep native/predicted coordinate provenance separate from experimental validation. Optional vdjdb build --structure-supplemental <metadata.tsv.gz> merges their raw metadata within one build. Leave it off in default hosted assembly/scoring until matched-cohort extraction is validated. CI uploads these reports separately from the release bundles. See structure cross-links.

For record confidence v3, complete motifs and run vdjdb score-records --tables out/tables with the confidence extra. The frozen model uses accepted predicted-structure metrics; native coordinates remain separate evidence. Parameters stay frozen as available evidence changes; held-out assay metrics use the complete original structural inputs. Default membership is hard; exploratory --soft-membership and its batched search/count fields are documented in the specification. See record confidence v3 for its precise inputs and outputs.