Responsible use.
What Tulparus will and will not do, written down.
What it is
Tulparus is decision support for clinical genetics teams working on cases that stayed unresolved after first-line testing. It is not a diagnostic system.
It assembles a case, ranks hypotheses and shows the evidence for and against each one, proposes next steps, drafts documents from the record, and re-checks cases when the knowledge changes. Each step cites its source and records the versions that produced it.
What it will not do
- Give a diagnosis, or show a rank or a similarity score as the chance that a gene or disease is the answer for a patient.
- Approve, dismiss or sign anything. Those steps belong to people.
- Treat a feature nobody examined as absent. Present, absent, unknown and not tested stay separate.
- Mix candidate new genes, which have little or no established evidence, with established findings.
- Overwrite conflicting evidence. Both sides stay on the record.
- Use case data to train a model without the organisation’s recorded agreement.
Human review
Every proposal is a proposal. A person reviews it, accepts, changes or dismisses it, and the reason is kept. Reports are rendered from the approved record. We ask users to treat what Tulparus shows as a place to start looking, and to check it against the source before acting.
Limits
Tulparus can only work with what it is given and what the knowledge release contains. In particular:
- A poor sample, an incomplete phenotype or a missing family member limits what it can find. It tells you what is missing instead of guessing.
- Knowledge changes. A result is tied to the release that produced it and may look different on a later one.
- A candidate new gene is a lead with little evidence. It is not a finding.
- A benchmark hit rate describes a model release on a set of published cases. It says nothing about the chance for the patient in front of you.
How we test
Models are evaluated on fixed sets of published cases, with the knowledge limited to what was available when each case was reported. We write down the protocol before a run, and report the result with the size of the set and a 95% interval. A new model does not replace the current one because its training loss is lower. See Validation and Research.
How to raise a concern
If a proposal looks wrong, dismiss it with a reason. The dismissal is recorded and we review it. For anything else, such as a possible harm, a wrong claim on this site or a security problem, write to [email protected]. Security reports go to [email protected].