HLA-Verify

Every HLA allele name, checked against the source. Deterministically.

AI assistants are entering transplant and immunogenetics workflows — and they fabricate allele names, resurrect deleted ones, and slip into 1990s formats. HLA-Verify checks every allele-shaped token in a typing report, EHR fragment, or model output against a pinned IPD-IMGT/HLA release. No LLM anywhere in the loop: every verdict is computed, versioned, and reproducible.

Patient typing: A*0101 → A*01:01 B*15:504:01 DRB1*14:06  Assistant suggested: DQB1*05:03:26:99 ✗ fabricated

Fabrication caught

A 7B open model fabricates allele names in 9% of benchmark tasks and answers 62% of them wrongly with high confidence. Our benchmark, HLA-Bench-A, measures it; this engine catches it at runtime.

Versioned to the release

Pinned to an exact IPD-IMGT/HLA release (46,652 alleles in 3.65.0, ~600 added per quarter). Deleted names resolve to their successors across 110 releases of history.

Nothing stored

Reports are processed in memory and discarded; we log counts, not content. Reference data are fetched from the official source under CC-BY-ND and never redistributed.

Hosted API — live at api.hlaverify.com

Four endpoints, JSON in and out, pinned to IPD-IMGT/HLA 3.65.0, served from Cloudflare's edge in well under a second. /v1/verify checks free text, /v1/normalize fixes typing strings from any era, /v1/allele returns the facts, /v1/match scores a donor–recipient pair with the published rules (per-chromosome counts, GvH/HvG, null-allele traps, and potential when the typing is too coarse to call). Open for evaluation at 60 requests/minute; keyed access for labs, LIMS vendors and agent platforms.

curl -s https://api.hlaverify.com/v1/verify -H 'content-type: application/json' \
  -d '{"text": "Patient typing: A*0101, B*15:504:01. Assistant suggested DQB1*05:03:26:99"}'
→ {"clean": false, "counts": {"valid": 1, "deleted": 1, "hallucinated": 1, …}, "tokens": [{"token": "DQB1*05:03:26:99", "status": "hallucinated", …}]}

Free — demo & open source

In-browser verifier (nothing leaves your machine), open benchmark, Python engine on GitHub. Apache-2.0 for the benchmark and graders.

Pilot — $12k, 6 weeks

Your reports or model outputs, verified against a pinned release with a full audit trail and a findings letter. Fee credited toward a first-year licence.

Licences — from $15k/yr

Commercial licence for the verification service on your infrastructure; model-evaluation licences and custom environment families for AI labs — talk to us.

Seeking partners: registries and labs holding licensed data can run population-realistic evaluation slices on their own infrastructure — your data never leaves. AI agents: see /llms.txt.

Try it now — runs in your browser API docs Get an API key Benchmark & code

The demo runs entirely client-side — nothing you paste leaves your machine; the hosted API stores nothing and meters counts, never content. Keyed API for labs and LIMS vendors: pilot slots open.