Performance measured.Limitations made clear.

Tulparus evaluates its models on held-out clinical cases with documented benchmarks and reproducible evaluation methods. Results include dataset details, sample sizes, performance metrics and uncertainty intervals, showing where the models succeed, where they fall short and what the results mean.

Unseen cases.Knowledge frozen in time.

Tulparus evaluates its models on cases excluded from training and tuning. For new-gene benchmarks, scientific knowledge is frozen at a historical cut-off, so the models are tested against gene–disease links that were not yet established. Every release is assessed against documented benchmarks and predefined quality criteria.

Independent evaluation

Tested beyond the training data. Evaluation cases are held out from model development, with datasets, sample sizes and evaluation methods documented.

Time-frozen knowledge

Evaluated against what was known then. Historical knowledge cut-offs measure performance on gene–disease links not yet established at the time.

Release validation

Measured before deployment. Each model release goes through defined performance, reliability and regression checks, with results and limits recorded.

Benchmark performance describes an evaluated model on a defined dataset, not the probability of a diagnosis for an individual patient.

Different models.Measurable results.

Each Tulparus model is evaluated on its own across clinical scenarios. We report performance and limitations openly, so the capabilities of each model are clear.

Phenotype rankerMeasures how well clinical findings identify the relevant genetic candidates.
Genotype rankerEvaluates genetic prioritisation across different clinical scenarios.
New-gene rankerMeasures the ability to find promising candidates beyond established disease associations.
Note readerEvaluates the accuracy and reliability of clinical information extraction.
LimitationsPerformance varies with the evidence available, case complexity and data quality. Known limitations are documented alongside the results.

Ask for the details.

We share the current model card and benchmark card with qualified teams, including the knowledge releaseA dated snapshot of the reference databases and literature the models read from, so every result can name what it was based on. used for each run.