# Variation is not noise

10 December 2025 · Essay · 7 min read

By [Ahsan Fazal](https://axiomatic.digital/en/about#ahsan), Founder and CEO

The case for replacing professional judgement with algorithms stands or falls with one assumption: that a brain is a computer. Where a judgement has no right answer apart from the judgement itself, difference between professionals is information. Give them the same sources, not the same conclusion.

The argument someone forwarded to me overnight goes like this: professionals who judge the same case reach different conclusions. That difference is noise, a flaw in the system. An algorithm given the same input produces the same output every time. So human judgement is better left to algorithms.

This morning I wrote back that the reasoning holds together on its own terms, but that I disagree strongly with its conclusion.

Two things first. I have not read the book the argument leans on, so I am answering the argument as it was put to me, not its author. And I know too little about behavioural science to judge that field. What I do know something about is what a system does with differences between professionals. That is what I build.

For claims far outside someone’s own field, my standing test is to take the contrary position and let the arguments convince me. I did. I am not convinced.

## One assumption

The argument stands or falls with one assumption: that a brain is a computer. Accept that assumption and the logic follows. Reject it and it does not.

Follow the logic through. If a brain is a computer, a judgement is a calculation. The same input should then give one right output. If two nurses reach different answers to the same question, at least one of them has miscalculated. Variation is then an error by definition, and a machine that calculates the same way every time is better by definition. Each step follows from the one before. Everything hangs on the first.

I do not take that first step. I work with language models every day. The parallels between how people learn and how such a model is trained are striking, and I am not immune to them. Eight days ago, after a lecture on how language models are trained, I wrote: “tacit knowledge (human) = tacit ‘experience’ (LLM)”. An equals sign. I meant an analogy, and that is how the sentence began. But that is how fast it goes. You see a resemblance, you write an equals sign, and three steps later you have replaced a nurse’s judgement with a formula.

This morning I put it more precisely. Everyone keeps being astonished by the striking parallels between how people learn and think and how language models do, but no one would claim they are the same: the differences are too great, and we know too little. Resembling something is not being it. Jumping from one to the other is a misstep in thinking. It is an assumption dressed up as a result.

## What the word noise assumes

The word itself gives the assumption away.

Noise is deviation from a signal. To speak of noise is to say there is a signal: one answer that is right, whoever gives it. For part of the work in care, that holds. Which version of a protocol applies today. Whether a document has expired. Which source the organisation gives precedence to. What the work agreement says. A sum. There, a right answer exists outside the professional’s head, and anyone who deviates from it is wrong. That variation is an error, and it has to go.

But a large part of the work has no right answer apart from the judgement itself. How do you give hard news to a family that is angry? How do you write a handover so that the night shift reads what matters? How do you explain to a new colleague why the ward does something differently from what her training taught her? How do you tell a colleague that an agreement in the team is not working? Two experienced professionals do these things differently, and both can be right. The difference tells you something: about the family, the moment, the team, the person doing it.

There, variation is not noise but information.

Filter that variation out and you keep an average and throw the signal away.

## Which error you want to reduce

The question I think it comes down to: how is the task of judging actually built, and which error are we trying to reduce?

Take a judgement apart and you usually find two parts. What you need to know: the facts, the source, the version, the agreement. And what you need to weigh: what counts most here, how you say it, when. The first part has a right answer. The second often does not.

The big error sits in the first part, and it rarely sits in the professional’s head. It sits in what that professional can know. One reads the current protocol, the other a printout two years old. One walks into an experienced colleague’s office during the day, the other is on her own at night. They reach different answers, and that difference really is noise. But the remedy is not an algorithm that decides for both of them. The remedy is that both of them read the same thing.

This is where the reasoning forks. Call variation noise and you give everyone the same conclusion. Take the task apart and you give everyone the same sources.

## What that means for what we build

**The same sources.** We build Inora to answer from the organisation’s own protocols, work agreements and quality handbook, with the source alongside, day and night. The same question should get the same answer, even where two sources disagree: the same order always applies. What is settled is settled for everyone, for the night shift as much as for the day shift.

**The decision stays with the person who answers for it.** Inora decides nothing about a client, nor which source leads, nor which method its answers follow; the professional and the organisation do. It answers about sources and about what the care worker wrote, and it judges no client. When a question holds a decision about a client, Inora should answer the part a document answers and say who decides the rest, under the organisation’s own protocol. A nurse who knows what the protocol asks still decides for herself. Giving knowledge is not taking a decision.

**The steps in view.** Where the work asks for thought, Inora should set out the steps the organisation’s method prescribes, each resting on a source, and record them, so anyone can check how an answer came about. The conclusion stays with the professional.

That last part makes variation readable. When two colleagues reach different answers, you can see which kind of difference it is. If they read different things, it is an error, and you fix it at its cause: in the source or in the access. If they read the same thing and weighed it differently, it is information. That is something for the team to talk about, and sometimes it turns out the quality handbook is missing a sentence.

And we do not line care workers up against each other to measure who deviates: Inora is not there to assess staff. A system that counts differences between colleagues as errors mostly teaches a team to write down the same thing. Then the difference disappears from the record, not from the work.

## Where the argument is right

The argument is not worthless because its assumption fails. Quite the opposite.

Where a right answer exists, it is right. People are bad at doing exactly the same thing twice, and code is good at it. A sum, a date, a version number, whether a document still applies: there is no room for variation there, and that is work for a machine. The mistake is stretching that point to cover every judgement. Then an argument for better sources becomes an argument against the professional.

## The same answer makes difference invisible

I still know too little about behavioural science to judge the field the argument comes from. What I do know: a system that gives everyone the same conclusion does not make the difference between professionals smaller. It makes it invisible, and with it whatever you could have learned from it.

Give them the same sources. Leave the judgement with them. Record how they got there. Then a difference between two professionals is not noise to filter out, but a question a team can ask.

*The quotation is translated from Dutch; its English terms are kept as written.*

## Work in care and use Inora?

Sign-in and help go through your own organisation: your team’s project lead and ambassadors can help you.

## Read on

- [Rebuilt on open building blocks](https://axiomatic.digital/en/updates/rebuilt-on-open-building-blocks.md): Six weeks in, we are rebuilding Inora from the ground up on open-source building blocks. A small company does not win by building everything itself, or by letting one platform make its choices. What we take, what stays ours, and what it costs.
- [A language model does not reason](https://axiomatic.digital/en/updates/a-language-model-does-not-reason.md): In 2025 I announced “reasoning” as a feature. The word was wrong. A language model picks the most likely next text, astonishingly well, and does not know what it does not know. So: the model writes, the sources decide what is true, code does what must be exact, the professional decides.
- [An evaluation set makes “any model” true](https://axiomatic.digital/en/updates/an-evaluation-set-makes-any-model-true.md): Since the rebuild I say Inora is tied to no single language model. The sentence is only true if we can show it. Without the organisation’s own questions and answers, swapping a model is a leap of faith; with them it is a test run. Why I count evaluation as a precondition, and where I part ways with the standard advice.