# Assembly-derived reference labels

Start with [haplotype_assignments.tsv](haplotype_assignments.tsv). Each row describes one donor × gene × archived chromosome copy, not an entire multigene MHC haplotype.

| Column | Meaning |
| --- | --- |
| `donor`, `resource`, `gene` | Source donor, cohort and HLA locus. |
| `haplotype` | Archived haplotype label. Preserve it when joining to retained input sequences. |
| `public_haplotype` | Source-publication haplotype label after the explicit Saudi crosswalk. KSA006 and KSA009 require reversal. |
| `basis` | Genomic, CDS, protein or unresolved sequence evidence. Inspect this before choosing an evaluation endpoint. |
| `genomic_matches`, `coding_matches`, `protein_matches` | All matching allele labels at that evidence level. Empty means no evidence of that type; it is not a predicted homozygote or absent gene. |
| `numeric_two_field`, `numeric_three_field` | Supported unique numeric prefixes where the generator can resolve them. These do not add missing fields or infer expression suffixes. |

For full evidence use `DIRECT_MATCHES.json` (HPRC genomic matching), `CODING_MATCHES.json` (HPRC CDS/protein assessment) and `EXTRA_MATCHES.json` (Arab/Saudi retained intervals). `COMPLETE.json` records checksums and semantic reproduction. The TSV is a convenient projection of those records.

Choose a scoring resolution explicitly. A lower-resolution rescue does not establish genomic/fourth-field identity. Keep unresolved or ambiguous cells explicit; do not select an arbitrary matching name. A diploid endpoint requires eligible evidence for both copies. Do not silently replace reference labels with a caller's output. Published method scores have their own endpoint eligibility and denominators.

These are the reference-label candidates for this developmental benchmark. `final_truth_admission=false` preserves the outstanding read support, phasing, quality, relatedness and exposure assessments. KSA001 accession-to-haplotype association remains unresolved beyond the graph binding. Whole-MHC phase correspondence is not inferred from gene-level sequence equality.

[Release guide](../README.md) | [Supporting inputs](../inputs/README.md)