Let your scientists set the rules your lab's AI follows

To set up an AI agent for your lab, start with your own scientists. A screening group that calls a compound a hit when it clears the group's cut-off on both replicate plates with passing controls has already decided which evidence counts and what makes a signal real. Elnora, the AI agent for lab work, takes those rules from the scientists who set them and applies them from the first task.

The judgement an agent needs already sits with your scientists

A general method is a starting point: a literature review, a dose-response fit, a safety summary for a target. The version a lab acts on reads the lab's in-house data, such as a toxicology dataset that stays inside the building, and scores it against thresholds the group argued its way to over years of assays.

Fitting an agent to a lab means taking positions: which evidence sources count for this decision, what effect size is real for this assay, which internal dataset overrides the public one, and what a safety assessment has to contain before the review committee accepts it. The principal scientist who set the hit cut-off can defend each of those in front of colleagues, and the group head knows what the committee asks to see. Some vendors answer this by sending their own PhD scientists to sit with each customer and tune the agent. Your lab already employs the people who hold those positions. Elnora is built for them to set the rules directly, and a change they make applies from the next task.

A scientist gives Elnora the lab's judgement once

Elnora plans experiments, writes and optimises protocols and analyses results, and it keeps a lab's standing decisions in memory. The group's conventions sit in one plain file in its workspace, open to colleagues and to the agent for editing, and Elnora loads that file into each turn of a task. The lab's own results, protocols and instrument exports go into the Knowledge Base, which Elnora reads before it writes anything.

Take the screening scientist. Before the first analysis they tell Elnora the group's rules: a plate counts when its Z prime passes the group's threshold, a compound is a hit when it clears the cut-off on both replicate plates, and the group's own counter-screen results outrank a published activity value for the same compound. Elnora writes those rules into the team's conventions.

The next week a colleague uploads a new batch of plate reader exports and asks for the hits. Elnora decodes the exports into tidy tables, runs Z prime QC and marks the plate whose controls sit too close together, applies the cut-off across the replicate plates, and checks each hit against the counter-screen results in the Knowledge Base before it searches the literature. The report carries the record of how it was made, and the literature on each hit closes with its source list.

When the group moves its cut-off, the scientist edits the conventions file and the next analysis uses the new value. If the same analysis is run three times, Elnora saves it into the scientist's own toolbox as a named skill, a written method it follows the next time, so the method and the rules it follows are kept together.

The judgement goes from the scientist who holds it straight into the agent, and a change to it applies from the next task. That is the mission we build Elnora for: every scientist codes their experiments, and Elnora does the lab work by that scientist's rules.