Free API key, instantly: 1,000 calls a day. Get a free key, or use it with no key at 100 a day.
HLA-Verify

Trust

What we do with your data. Which is nothing.

Verification is a job you should be able to hand to a service without adding a data-governance problem. This page is the whole posture — data handling, reference-data licensing, reproducibility, regulatory scope and code licence — written plainly enough to forward to whoever has to approve it.

The two documents a reviewer usually asks for next: the privacy policy, which lists every field the service stores, and the terms of use, which say what the free beta is offered on.

Your input

  • Not stored. Request bodies are processed in memory and discarded when the response is written. There is no request log of content, no queue, no analytics capture of what you sent.
  • Counts, not content. Operational metrics are call counts and token counts per key. That is what an audit log contains, and it is enough to bill and to rate-limit.
  • Not used for training. There is nothing to train — the product has no model. Your input is not training data because no training happens.
  • Not shared. Nothing you send to the API reaches an analytics vendor, an advertising pixel, or any external call made on your behalf during a request. (This site, separately from the API, measures page views with Cloudflare Web Analytics, which is cookie-free and never sees a request body. Privacy says what it collects.)
  • Self-hosting available. If the answer to "may data leave the network" is no, run the container yourself and the question does not arise.
On identifiability, stated honestly. HLA-Verify is designed to take allele names, not patient records — and it never asks for identifiers. But HLA types are themselves genetic data, and a rare type in a small population can be quasi-identifying even without a name attached. So treat what you send accordingly, send names rather than whole records where you can, and if your policy requires that such data stay inside your perimeter, self-host. We would rather say this than let it go unsaid.

Reference data

HLA-Verify computes from IPD-IMGT/HLA. It does not contain, republish or resell it.

  • Fetched from the source, not bundled. Release files are downloaded from the official ANHIG/IMGTHLA distribution and MD5-verified before anything is computed from them. The self-hosted container does that at startup and caches the files locally. The hosted API does it at build time instead: the release is loaded once, the same engine that grades the benchmark computes the lookup tables, and those derived tables ship with the deployment, which is why the hosted service answers without reaching out to anyone mid-request. Either way the database itself is not redistributed in our container, our repository or our responses, and the bulk tables are not served.
  • Integrity-checked. Every fetched file is MD5-verified against the release's own md5checksum.txt before it is used. A mismatch is a hard failure, not a warning.
  • Attributed in every response. Each API response carries the source attribution and the release it was computed against.
  • CC-BY-ND respected. The database is published under CC-BY-ND. We compute over it and cite it; we do not distribute modified copies of it.
  • Independent. HLA-Verify is not affiliated with, endorsed by, or acting on behalf of ANHIG, IPD-IMGT/HLA, WMDA or NMDP.

Citation: Barker DJ et al. The IPD-IMGT/HLA Database. Nucleic Acids Research, 2025.

Reproducibility

The claim is specific: the same input, against the same pinned release, produces the same output — always.

No sampling

No temperature, no seed, no sampling step anywhere. The code path from input to verdict is a lookup and a set of fixed rules.

No silent updates

One release is compiled into the service and every caller gets it, named in every response. It does not move because we shipped something else: moving it is a deliberate rebuild, published with a diff of what changed. Holding an individual key on an older release is planned and not built, so today the honest version is that the release is the same for everyone and always visible.

Verifiable months later

Every response names its release. Given the release, any verdict can be recomputed and checked — during a study review, an audit, or a disagreement.

Regulatory posture

HLA-Verify is not a medical device and is not offered as one. It performs nomenclature verification against a public reference database. It does not perform HLA typing, does not interpret results, does not determine donor suitability, and does not produce a diagnosis, a treatment recommendation, or any clinical determination. Match counts and the compatibility endpoints are decision support only: they compute a published rule (the counting frameworks, the HLA-B leader rule, KIR ligand groups) over the typing you supplied, and they assume that typing is correct. Reading those outputs against a patient is the clinician's judgement, not ours.

Its output supports, and does not replace, clinical judgment and your laboratory's own validated procedures. Responsibility for validating any tool within your accredited workflow — including this one — remains with your laboratory, as it does for every other tool you use.

We would rather be plainly useful inside that boundary than vague about where the boundary is. If your accreditation body asks what this thing is, the answer is: a reference-lookup service that tells you whether a name is real, current and correctly formatted, and shows its work.

Licensing

Open-core, deliberately. The evidence is open so it can be checked; the service is licensed so it cannot simply be resold.

ComponentLicenceWhat that means
Benchmark, task generators, graders, harness, framework adaptersApache-2.0Use it, fork it, build on it, commercially. The evidence has to be checkable to be worth anything.
The verification service (sci_envs/service)PolyForm Noncommercial 1.0.0Read it, evaluate it, use it noncommercially. Commercial use requires a licence — that is the product.
IPD-IMGT/HLA reference dataCC-BY-ND (upstream)Fetched at runtime, attributed, never redistributed.

The sealed benchmark split, the regeneration cadence and commercial agreements are not published. Everything needed to check a claim on this site is.