OmevraOmevra · About
About this project
Five working systems, one rule — cite the evidence or decline — and the reasons behind them, in plain English.
What was built
Top ↑Omevra is a set of five working systems built around one problem: a genetic variant whose significance is uncertain, and the evidence that bears on it. Each one is live, and each one publishes its own measurements.
Takes a variant of uncertain significance and writes an evidence summary in which every claim cites a retrieved record — a calibrated probability where the evidence supports one, and an explicit decline where it does not.
An autonomous agent that operates ordinary web interfaces — forms, tables, dialogs, portals — checking each step and recovering when one fails. It is measured on a fixed 36-task suite, run with its recovery mechanisms on and off.
A whole-genome sequencing pipeline for bacterial data, run end to end on real sequencing runs — then run a second time to prove the results reproduce exactly.
A benchmark for predicting how single cells respond to genetic perturbations. Baselines are published first, and the project’s own hybrid model is reported where it lands — including below the simplest baseline.
Three cooperating agents — one retrieves the evidence, one drafts the summary, one checks every citation — measured head to head against a single agent doing the same task in one pass.
Why it matters
Top ↑Sequencing a genome is now fast and cheap; classifying what it finds is not. A large share of the variants a laboratory reports come back as variants of uncertain significance — VUS — because the public record holds too little evidence, or evidence that conflicts between submitters, to support a call.
A tool that answers anyway, on thin evidence, makes that worse. So every system here follows the same rule: show the evidence behind a claim, show what is missing, and stop where the evidence stops. The measurements — accuracy, calibration, coverage, failures — are published with the systems, so a reviewer can check the claims instead of taking them on trust.
What comes next
Top ↑Plans, stated as plans — none of this is built yet:
The pipeline method applied to a human genome: the GIAB HG002 reference sample, scored with hap.py precision and recall for SNVs and indels — the standard clinical-genomics yardstick, beyond today’s bacterial demonstration.
Splice prediction and population allele frequency as evidence, so the Variant Evidence Agent can score more of the variants it currently declines.
A protein-structure view with the variant’s residue marked, so a reviewer can see where in the molecule a change sits.
A short preprint of the temporal-split calibration result — the finding that calibration fitted on past classifications can fail on future ones, and what fixed it.