Why
A negative report is not an answer.
We started from what happens to a case after the exome or genome comes back without a diagnosis.
The odysseyIn rare disease a family can wait years for a name for what their child has. They pass through clinics, repeat tests, and are often told at the end that sequencing found nothing that could be used.
What negative meansUsually it means that no variant met the bar on the day it was analysed. It does not mean there is no cause. Genes get linked to diseases, databases change, and a clinic note mentions a feature that nobody entered.
Where the time goesNot in the sequencing. It goes in re-reading the case: the same lookups, in the same tabs, by the same few people, between other work. The reasoning ends up in someone’s head and the next person starts again.
The work
What a genetics team does with a case that stayed open.
We have written down what we saw as a job to be done, not as a pitch.
The list is where it startsA report arrives with a list of variants. Someone filters it, then opens a dozen sources: population frequency, gene and disease evidence, earlier classifications, papers, the pedigree. Then the whole thing is repeated for the next case.
Three in the afternoonThe phenotype is in a clinic letter. The family history is on a scanned page or a drawing. The geneticist has twenty minutes before the next patient. What gets assembled in that time is what the case gets.
Why a workbenchA workbench is where work is left half done and picked up again. The case keeps its own state: the open questions, the tests ordered, the results we are waiting for. It does not produce a verdict and move on.
What we hold to
What we will not do.
These came before any model did. The product is built around them.
Give a probability of diagnosisA model can rank candidates, and a rank tells you where to look first. A number beside a gene gets read as the chance that it is the answer for this patient, and it is not. So we show the reasons: the evidence, its source, what is missing. When we report how a model performed, it is a benchmark result with its evaluation set and sample size, labelled as that.
Treat unknown as absentA feature nobody examined is unknown. It never counts against a hypothesis. Present, absent, unknown and not tested are four different values.
Hide where a claim came fromA claim, its source, the release and the record id travel together.
Let a model signModels and agents propose. Approvals and signatures belong to people.
Overwrite what disagreesIf two sources conflict, both stay on the record and the conflict is shown.
Make a diagnosis for youTulparus is decision support for cases that stayed unresolved. The clinician decides.