A ranked list with no sense of its own reliability is a way to launder a guess as a recommendation. Petitri is our attempt to do the opposite.
What it does
Petitri is a drug-discovery preview built on the Tilelli Med knowledge-graph embeddings. Point it at a disease and it ranks candidate drugs by how strongly the graph supports a link — repurposing hypotheses, in plain terms. It runs on the same compact ternary substrate as the rest of the Tilelli line.
The confidence dial
The part we care about most is the per-query confidence. For some diseases the graph has dense, mutually reinforcing evidence and the ranking is worth taking seriously. For others it's sparse, and Petitri says so rather than dressing up a thin signal as a strong one. Same instinct as the chat model's "I don't know," applied to a ranked list instead of a sentence.
Where the moat is
How the confidence is derived — what makes it trustworthy rather than a softmax in a trenchcoat — is the interesting research, and it's not in this post. We'd rather show the behavior than over-explain the mechanism before it's locked.
The disclaimer that isn't boilerplate
Petitri is a research tool for clinical review, not for treatment decisions. It surfaces hypotheses for experts to test. It does not prescribe, and a high rank is an invitation to investigate, never an instruction to act.
Want updates on the preview as it opens up? Tell us.