Reinventing the ELN for regulatory rigor

01Better science, not just better documentation

In their essay Getting Data to Grow in a Dish, insitro’s Daphne Koller and Max Salick describe eight years of building an automated experimental platform. Some of the most consequential errors they found were almost invisible: a mislabeled primer, a faulty USB port, a camera driver shifting images by a single frame. None necessarily raised an alarm, yet each could compromise the results. When experimental conditions are not captured, even sophisticated AI models struggle to distinguish real biology from errors introduced by the laboratory itself.

In regulated industries like pharmaceuticals and biotechnology, the stakes are higher still. Lab records must establish what happened, under what conditions, and why a particular decision was made. A regulator may need to reconstruct an experiment years later, down to the materials, instrument readings, and reasoning behind its conclusions.

Yet most electronic lab notebooks (ELNs) are little more than digital paper. Scientists work across instruments, spreadsheets, and analysis software, then assemble the documentation afterward. Missing controls, unrecorded conditions, and unexplained decisions often surface during review, when the original context may already be lost. The cost shows up as repeated experiments, delayed submissions, audit findings, and scientists’ hours spent reconstructing work instead of doing it.

What if the notebook could establish scientific rigor as experiments happen, instead of asking scientists to prove it afterward?

That is what we are building with Eureka at Frontier Instruments. Our AI notebook works alongside scientists, connected to protocols, instruments, previous experiments, and lab data. Before a run, it checks the planned experiment against the protocol and the lab’s past work to identify missing controls and metadata. During the experiment, it captures instrument readings directly and flags deviations. Afterward, it connects the results to the conditions and decisions that produced them, creating a record that colleagues, quality teams, and regulators can follow (Figure 1).

Crucially, this rigor does not come at the cost of more paperwork. Eureka drafts the record from the protocol and the run itself, and asks scientists only for what it cannot infer: the decisions and the reasons behind them. Documentation becomes a byproduct of doing the science, not a second job afterward.

02AI needs scientific guardrails

AI agents can already interpret experimental results, write analysis code, and propose new experiments in minutes. But their reliability remains a serious challenge. An agent that makes ten independent decisions with 90% accuracy at each step has only a 35% chance of getting all ten right (Figure 2).

OpenAI’s recently released BioBasicsBench makes this problem tangible. Across 100 routine wet-lab tasks, its best model earned just 66.7% of available credit at the highest reported reasoning setting. In one experiment, the model correctly grouped bands on a DNA gel but selected the wrong colonies because it ignored what the controls indicated. In another, it overcounted colonies by approximately 12%, directly distorting the estimated efficiency of the experiment.

We encountered the same class of problem in Eureka’s own evaluation on BixBench. On one question, Eureka reported 56.47% variance explained by the first principal component, narrowly outside the accepted range of 55–56%. The difference came down to an unstated choice about duplicate samples with conflicting metadata. Keeping, dropping, or combining them shifted the answer by 0.5 to 1 percentage point (Figure 3). The AI produced a plausible number without revealing that a consequential scientific judgment had been made.

In a regulated laboratory, this is precisely what cannot happen. So here is what Eureka does today when it meets the same situation. Before running the analysis, it detects the duplicate samples and their conflicting metadata, and pauses. It shows the scientist which samples are affected, the available options, and how each option changes the result. The scientist chooses, adds a one-line rationale, and Eureka records the choice, the alternatives considered, and the reason alongside the result. A reviewer opening the record a year later sees not only the number, but why it is that number.

The same principle applies wherever data conflicts, controls are missing, or instructions are ambiguous: the agent flags the uncertainty instead of quietly making assumptions, and scientists stay responsible for consequential decisions. Every record preserves what the AI proposed, what the scientist decided, and the evidence behind that decision.

The goal is not simply to make AI more accurate. It is to make the entire scientific process traceable, reproducible, and accountable.

03Built for the regulator in the room

In pharma and biotech, a lab record is only as valuable as its defensibility. Regulators expect electronic records to meet the FDA’s 21 CFR Part 11 requirements and the data integrity principles known as ALCOA+: every record attributable, legible, contemporaneous, original, and accurate, as well as complete, consistent, enduring, and available. Most ELNs treat these as compliance features bolted onto a document editor. Eureka treats them as the architecture.

Every entry in Eureka is attributed to a named person or to the AI agent, time-stamped when it happens, and preserved in an audit trail that cannot be silently edited. AI suggestions are recorded separately from human decisions, so a reviewer can always tell which is which. Approvals and sign-offs use electronic signatures linked to the record they approve (Figure 4).

This design also answers the question regulators are beginning to ask about AI in the lab: who made the decision? In Eureka, the answer is always a scientist. The AI proposes, flags, and drafts; the scientist decides and signs. That keeps regulated judgment where regulators expect it, while giving teams the speed of an AI collaborator.

04Regulatory rigor is the foundation for scientific intelligence

Here is the bigger opportunity: the information required to make experiments defensible to regulators is also the information needed to train reliable scientific AI.

Today, laboratories generate enormous amounts of data without preserving enough context to explain it. An IC50 measured at 1% DMSO after 24 hours is not equivalent to one measured at 0.1% after 72 hours. A cell model trained on microscopy images without magnification, exposure, or stain information may learn differences between experimental batches rather than biology. When those conditions are missing, models learn from results they cannot meaningfully compare.

Just as importantly, laboratories rarely preserve the experiments that failed. Inactive compounds, unsuccessful reactions, and screens that produced no meaningful result are routinely generated but seldom become structured data. This leaves AI models disproportionately trained on successes, making them overconfident about what will work in the real world.

Eureka addresses both problems by capturing the full experimental record: the design, conditions, raw measurements, negative results, deviations, and interpretations. But perhaps the most valuable information is the scientific judgment itself. Why did a scientist add a control? Why was a sample excluded? Why did they reject an AI-generated interpretation? These decisions contain the expertise of the organization, yet almost none of it is systematically recorded today.

By preserving the agent’s recommendations alongside the scientist’s corrections and rationale, every experiment can become training and evaluation data for the organization’s own scientific agents. Over time, the notebook accumulates not only what the lab discovered, but how its scientists reasoned their way there.

That asset belongs to the organization that created it. Each customer’s records, and any models trained on them, stay within that organization’s control; Eureka does not use one customer’s data to improve another’s. And because every model version is evaluated against the same audited record that trained it, organizations can show regulators exactly what their AI learned from and how it performs.

For R&D leaders, this changes the economics of the lab notebook. A traditional ELN is a cost center that grows with headcount. Eureka is an asset whose value compounds with every experiment: a proprietary dataset of conditions, outcomes, and expert judgment that competitors cannot buy and that makes each future experiment faster and more reliable (Figure 5).

05From system of record to system of scientific trust

This is why we believe the next generation of ELNs will be fundamentally different. A lab notebook should no longer be a passive archive of completed experiments. It should actively help scientists conduct rigorous work, preserve the evidence needed for regulatory review, and turn the resulting knowledge into an asset that compounds with every experiment.

We are reinventing the ELN from a system of record into a system of scientific trust, and ultimately, scientific intelligence.