The accessible human surfaceome

Benchmark, triage, and deep-dive data outputs for the accessible-surfaceome project. This draft ships three data files plus an in-deposit README documenting every column and the source-join recipe.triage-runs-genome-with-reasoning.tsv — 21,950-row long-format TSV consolidating the genome-wide Sonnet+NCBI sweep (19,324 genes × 1 rep, the ~19k-gene M1 candidate universe) AND a targeted PubMed-context re-run of a 2,626-row subset (the ambiguous-reason zero-DB Sonnet-no slice). A `run_id` column tags every row so the lanes can be split or merged at read time; for the ~2,621 re-run genes that's two rows (one NCBI, one PubMed). The read-side reconciliation rule (prefer PubMed when it is more inclusive than NCBI) is documented in the README and applied server-side by the live /v1/catalog endpoint.triage-benchmark-with-reasoning.tsv — 4,851-row long-format multi-replicate TSV covering the 147-gene curated benchmark across 4 models with uneven prompt-variant coverage (Haiku 4.5 + Sonnet 4.6: all 4 variants; Opus 4.8: naive + ncbi; Sonnet 5: ncbi) — 11 (model × variant) cells × 3 replicates × 147 genes — with curated truth verdict / signal / reason / class joined onto every row.deep_dives_all.tar.gz — one JSON per published SurfaceomeRecord (5,130 records), each carrying its full evidence chain, per-claim verbatim quotes, and the deep-dive classification (`deep_dive_tier` + `deep_dive_facet`) computed by the same predicate the viewer ships and attached server-side on /v1/genes/{symbol}.The manuscript will be added to this record in a later draft update against the same reserved DOI (10.5281/zenodo.20805384). All data files are reproducible from the public read-only API at https://api.deliverome.org/surfaceome/v1/ (no credentials).