Keep the runs that failed with the protocols that produced them

A knowledge base for lab protocols and results is the material a lab's predictions are made from. World models, systems that learn how an environment behaves and use that description to predict what happens after an action, arrived in medical research this year, and the question they leave a working lab is a practical one: which descriptions of your system have kept predicting the next result, and where does the record of those descriptions sit.

What a world model is

The term comes from robotics and from software that learns to control machines and play games, work that grew through the last decade. Acting on a real robot or in a real process is slow and sometimes expensive, so researchers taught systems to learn a model of their environment and try planned actions inside the model before acting in the world. A world model is a working description of how an environment behaves, small enough to keep and to check, accurate enough to predict what happens next and to choose between actions.

The route you drive to work is a small working case of the idea. You tried a few routes, you kept the one that avoided the level crossing at half past eight, and the description you now hold is a short sequence of turns and one time to stay clear of, kept because of what happened on the days you tried the alternatives. Your phone's next-word suggestion is built the same way. The suggested word comes from the words typed immediately before it and from patterns in a very large body of earlier text, and the description behind the suggestion is far smaller than a grammar of the language. Both descriptions are small, both were kept because they kept being useful, and both are consulted before the next move.

What biology does with small descriptions

Living systems ran on small descriptions long before software did. A honeybee returns to a feeding site across a landscape using the angle of polarised light in the sky, landmarks it has learned and the movement of the visual field across its eyes as it flies, and a description that small gets the animal to the site and home again. Prediction from a small record is also how a scientist reads a chromatogram and decides the next fraction worth collecting.

The regularities that decide a lab's next experiment are local ones. A knowledge base for lab protocols and results holds how a stock behaves after a certain number of passages, which buffer pairing failed in which assay, and which incubation step a team shortened and then came to rely on, and the version of those facts held in the lab's own record is the version that matched what its bench actually produced.

Descriptions earn a lasting place by predicting. A protocol step that has survived fifty runs is a short description of how the system behaves, with fifty runs of evidence behind it, and that evidence stayed in the lab, which is why a published paper cannot carry it. A failed run marks a description that stopped predicting, and filing the readout beside the protocol that produced it keeps the failure readable the next time those conditions come up. A record built from the successful runs supports confident predictions about conditions the failed runs would have warned against.

The clinical end of the field is moving quickly. One research system described in 2025 generates how a tumour's appearance would change under each candidate treatment plan and ranks the plans by simulated survival outcomes, and a review published in July 2026 frames the field as medical prediction moving past the static snapshot, toward models that represent how a patient's state changes over time and in response to what a clinician does. Clinical use lags the research, and a lab can run the same idea today on materials it already has.

What Elnora does with the lab's record

A scientist asks why this week's readout came back half of what the same assay gave a month ago. Elnora searches the knowledge base with the question in full and returns the passages that answer it, naming the file each passage came from, so the answer is assembled out of the lab's own protocols and results and each piece of it can be opened at its source. From the result in front of it, a skill for generating hypotheses drafts candidate explanations and ranks them by the evidence behind each, and the scientist picks which of those is worth the next attempt.

The protocol that follows carries whatever check that explanation calls for. Before any connected instrument executes the run, Elnora rehearses it in simulation and the scientist approves it, and once the readout returns, it is filed in the knowledge base beside the protocol that produced it, in the folder the pair belongs to, the day it lands. A run that fails lands there with its protocol, so a later question about those conditions returns both of them in the search.

The selection stays visible. Every protocol edit is saved as a separate version, each version recording its author and the version it came from, so a method that changed across weeks reads as an ordered record of decisions, showing what a team kept and what it replaced. What Elnora learns about how a team works persists the same way: notes carried between conversations are read back on every turn, the notes about how one person works stay private to that person, and the ones holding conventions the whole team depends on are shared across the organisation.

A lab that files each run beside the protocol that produced it, the failed runs included, has already written the working model of its own science. What an agent adds is reading that model before it writes, and filing what happens next back into it.