About
Built by someone who has had to get these names right
HLA-Verify comes out of immunogenetics and bioinformatics work — matching, imputation, haplotype frequencies, and the unglamorous business of reconciling allele names collected across decades and formats. The failure mode it addresses is not hypothetical.
The observation
Language models are arriving in transplant and immunogenetics workflows: drafting reports, summarizing typings, answering questions from laboratory staff. They are useful. They are also confidently wrong about HLA nomenclature in a way that is unusually hard to catch, because a fabricated allele name is syntactically perfect. DQB1*05:03:26:99 looks exactly like a real name. Nothing about it flags itself to a reviewer.
Meanwhile HLA is one of the few areas of biology where ground truth is genuinely public, authoritative and computable. IPD-IMGT/HLA publishes the assignments, the deletions, the renames and the group memberships, release by release, back to 1.05.0. If the truth is a lookup, then checking should be a lookup too — not a second model's opinion about the first model's output.
Two products, one engine
HLA-Bench — measuring the problem
A benchmark of deterministically generated, deterministically graded tasks that quantifies how badly models handle HLA nomenclature, and which failure modes they exhibit. Published in the open, regenerated each release, with a sealed split so scores mean something. See the results →
HLA-Verify — fixing it at runtime
The same graders, run forward as an API. What scores a benchmark answer is precisely what should check a report before sign-out, or gate an LLM feature before its output reaches a laboratory. See how it works →
This is not two projects that happen to share a domain. The grading engine is the verification engine. Every task family added to the benchmark becomes a capability in the API, and every edge case found while grading makes the API more correct.
Who
HLA-Verify is built by Jason Brelsford — a bioinformatics and infrastructure engineer working in HLA immunogenetics, with a peer-reviewed publication on tooling in this domain and hands-on work across matching, imputation, haplotype frequency data and registry-scale pipelines.
It is an independent effort, currently pre-revenue and in private preview. That is stated plainly because the people evaluating it are laboratories and vendors who need to know exactly who they would be depending on. What that buys you at this stage: the person who answers your email is the person who wrote the resolution rules, and early pilots genuinely shape what gets built next.
Principles
- Never generate ground truth. Every fact is computed from a versioned reference release or it is not reported at all. There is no model inside the product and there is not going to be one.
- Publish the evidence. The benchmark, the graders and the transcripts are open. Claims on this site are checkable, including the unflattering ones.
- Say what it is not. It is not a medical device, not a typing tool, not a matching algorithm, and not affiliated with the database maintainers. Overclaiming in a clinical domain is how tools lose the right to be trusted.
- Evaluate models, never compete with them. The position is verification. That is a durable place to stand precisely because it does not depend on winning a modelling race.