NILS
The registry our imaging lives in.
NILS keeps a neuroimaging group's data as one registry: every scan digested and classified by a rule pack, every person pseudonymous, every question a stored document with a version chain and a content hash, every release described in full before it is made. The desk is the one place to ask, review and release; an assistant beside it proposes and never decides. It runs the same on a laptop with nothing beside the one binary as it does on our own servers behind the portal.
Ask
A question starts from anything, a cohort, a saved selection, a result, a pasted list, and is built by clicking or by talking to the assistant. Its counts move as it is built, its declaration stays on the page, and its answer is a result that is paged, never run again.
Review and release
What the pack was not sure of waits in a queue sorted by what it costs if wrong, with the evidence beside it. A release names its pseudonym namespace, its layout, what it holds back and where it goes, and writes an audit row for everything it does.
Open source
We wrote NILS ourselves and publish it under the AGPL: a Rust engine, a Rust and React desk, a TypeScript assistant, and Kvasir, the model gateway. The record of every decision, the specifications and the gate are at github.com/kineuro/nils.