# T1K predictions — 2026-10-20

These are method predictions to compare against the [benchmark reference labels](../../../releases/2026-10-20/truth/README.md), not ground truth.

- [Download the complete prediction package](t1k-baselines-20261020.tar.gz) and [verify its checksum](t1k-baselines-20261020.sha256).
- [Scores and per-cell comparisons](track-a-scoring-20261018/TRACK_A_RESULT.json).
- [Native genotype tables and completion records](track-a-scoring-20261018/collected/).
- [File manifest and coverage](T1K_BASELINES_MANIFEST.json).
- [Input verification](interpretation-20261020/TRACK_INPUT_VERIFICATION.json) and [panel provenance audit](interpretation-20261020/CORE_PROVENANCE_AUDIT.json).

Coverage: 86 donors in each of 11 active reference configurations, 946 genotype tables. These are configurations of T1K, not 11 independent callers. The obsolete 60-donor original non-Asian configuration is excluded from exported genotype tables; retained historical aggregate/audit rows identify it separately.

The `genomic` configuration uses full genomic IPD 3.65. Flanked panel augmentations and deduplicated core augmentations represent different reference designs. The `hprc_asian` panels include Asian HPRC plus Arab and Saudi content. Read the recorded eligible endpoints and provenance before attributing a contrast specifically to ancestry or graph structure.

Predictions use the developmental assembly-derived reference labels; no untouched, independently read-certified test cohort is claimed. Missing/ambiguous truth cells require endpoint-specific treatment.

The archive preserves its original execution directory layout. Checksums verify the prediction payload. No raw reads are included.

[All prediction snapshots](../../README.md) | [Consortium guide](../../../README.md)