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, LIMS export or model output against a pinned IPD-IMGT/HLA release. No LLM anywhere in the loop: every verdict is computed, versioned, and reproducible from a release hash.
Free public beta: the API is open at 100 calls a day per IP with no signup, and verdicts are production-quality, pinned to IPD-IMGT/HLA 3.66.0. Paid tiers start at $49/mo and are set up by email; see pricing.
The problem
A wrong allele name does not look wrong
HLA nomenclature has 110 releases of history, colon-less legacy forms still circulating in the field, ~600 new alleles per quarter, and 289 names that were deleted or renamed. A fabricated name like DQB1*05:03:26:99 is perfectly well-formed. Nothing about it signals that no such allele has ever existed — not to a reviewer skimming a report, and certainly not to a language model that produced it.
of tasks contain a fabricated name
A 7B open model invents allele names in 47 of 550 benchmark tasks — invented fourth fields, made-up G/P groups, legacy forms.
wrong and confident
Its most common outcome is wrong_but_overconfident: 340 of 550 tasks answered incorrectly with no hedge and no abstention.
on deleted names
On the 100 tasks about names that were deleted or renamed — exactly where clinical risk concentrates — accuracy collapses.
Figures from HLA-Bench-A-v0.1@IMGT-3.65.0, ollama/qwen2.5:7b, all 550 tasks. Full tables, tiers, slices and per-subtype breakdowns are published.
The answer
Computed ground truth, not generated ground truth
HLA-Verify is a lookup engine over a pinned release of the official database, wrapped in an API. It has no model inside it, which is the whole point: it cannot hallucinate, it cannot drift between calls, and any verdict it returns can be reproduced from the release hash months later during an audit.
Fabrication caught
Every allele-shaped token is classified as valid, a G/P group, deleted (with its successor), a fabricated group, or hallucinated — no such name in any release back to 1.05.0. NMDP multiple allele codes are surfaced as mac_code, never silently skipped.
Versioned to the release
Pinned to an exact IPD-IMGT/HLA release — 47,424 alleles, 793 G groups and 1,580 P groups in 3.66.0. Every response names its release, and every move to a new one is published with the verdicts it changed.
Your requests are not stored
Verification requests are processed in memory and discarded. We log counts, not content. Reference data are fetched from the official source and never redistributed.
The API
One deterministic engine, every endpoint pinned to one release
Point it at free text and it returns a verdict per token. Point it at a column of typing strings and it returns current names, 2-field reductions, G groups and clinical flags. Ask about one allele and it returns the facts. The Lab Toolkit adds match counts, typing QC, compatibility models and GL String validation on the same engine: see the API reference.
$ curl -s https://api.hlaverify.com/v1/verify \
-H 'content-type: application/json' -H "x-api-key: $KEY" \
-d '{"text": "Donor A*0101, B*15:504:01; assistant read DQB1*05:03:26:99"}'
{
"release": "3.66.0",
"clean": false,
"counts": { "valid": 2, "deleted": 0, "group": 0, "fabricated_group": 0, "hallucinated": 1 },
"tokens": [
{ "token": "A*0101", "status": "valid",
"current_2field": "A*01:01", "flags": ["deprecated_name"] },
{ "token": "B*15:504:01", "status": "valid", "current_2field": "B*15:504" },
{ "token": "DQB1*05:03:26:99", "status": "hallucinated",
"note": "no such name in any release back to 1.05.0 \u2014 fabricated" }
]
}Who it is for
Wherever an allele name is about to be trusted
Accredited HLA laboratories
Roughly 200 ASHI-accredited labs in the US alone. As AI assistants enter reporting and review workflows, validating what they emit is becoming a compliance question, not a curiosity. Batch-normalize legacy typings while you are there.
LIMS & diagnostics vendors
You are embedding LLM features into products that laboratories sign out clinically. A deterministic guardrail on nomenclature is cheaper to build in than to explain afterwards. OEM terms available.
AI product & safety teams
A runtime check on life-sciences output, and a public benchmark to measure against. HLA is one of the few biology domains where ground truth is public, authoritative and computable — so the reward signal is real.
Registries & research groups
Harmonize typings collected across decades and formats to a single pinned release, with the mapping reproducible from a release hash for as long as the study is under review.
Start with the evidence, not the pitch
HLA-Bench-A is published in full — 550 tasks across 20 subtypes and 4 tiers, with the dev split, the grading rules and the model transcripts in the open. The same engine that grades the benchmark is the engine behind the API.