Field Note · Ontology

Your model can't beat your noise floor

15 June 2026 · Risto Anton Paarni · Helsinki, Finland

Most people picture an ML project as one thing: training. It isn't.

In reality a project is about 50% evaluation, 40% data cleaning, 8% integration, and 2% training. The part everyone imagines is the whole job is the smallest sliver of it.

Phase	Share	What it actually is
Evaluation	50%	Defining what "correct" means, building the test set, scoring honestly
Data cleaning	40%	Labels, dedup, reconciliation, removing contradictions
Integration	8%	Wiring the model into systems that act on the output
Training	2%	The part everyone pictures as the entire project

Read the top two rows again. Evaluation and data cleaning are ninety percent of the work — and they are exactly the part an off-the-shelf model hands you nothing for.

The floor is set before the model sees the data

Here is the part that decides everything. Evaluation and clean labels set the noise floor for learning. No model drops below it.

The model cannot lower the noise floor. That floor is the optimal bound of the Shannon encoding of your data — information your data never captured cannot be recovered by a bigger network.

This is not a tuning problem you can out-compute. Dirty labels are an information-theoretic ceiling. Buy a larger model, rent more GPUs, and you still hit the same wall — because the wall is in the data, not the math.

So the honest question is never "which model?" It is "what is our noise floor, and what moves it?" One thing moves it: a governed ontology that keeps labels clean, consistent, and current. That is the same thesis we keep returning to — the ontology, not the model, is the product — now stated in information theory rather than anecdote.

What a high noise floor looks like

It is rarely dramatic. It is old labels quietly going wrong, in every industry we work in.

Construction

A component tagged to last year's EN standard, a supplier cert that lapsed, an embodied-carbon factor nobody refreshed. The model learns the stale tag faithfully.

Energy & Manufacturing

Two systems disagree on a plant's emission factor and nobody reconciled them. Train on that and you have taught the model the contradiction, not the truth.

Agriculture & Logistics

A parcel's subsidy status or a consignment's customs class changed at the regulation boundary, but the label still reads the old value. Stale labels raise the floor for everything downstream.

None of these are fixed by training. All of them are fixed by reviewing old labels — continuously, as a system rule, not a spring-clean.

Why old labels are the job, not the chore

Not a single day goes by without thinking about ontology. Labels rot. Definitions drift. Yesterday's correct tag is today's contradiction. We built this conviction into the platform years ago and called it Active Pruning: old, contradictory, or invalidated knowledge is actively removed, and the system forces reconciliation between old facts and new stories.

We even set a target for it — a healthy 5–10% of labelled artifacts turn over every quarter. Turnover is not decay. It is the floor being held down on purpose.

The honest part

A clean ontology lowers the noise floor. It does not invent signal that was never collected. If the measurement was never taken, no governance and no model will conjure it — you go and collect it. And evaluation is the half nobody wants to fund: without an honest test set, you cannot even see your floor, let alone lower it. We will not sell label hygiene as a substitute for measuring the right things in the first place.

The short version

ML is ~50% evaluation, 40% data cleaning, 8% integration, 2% training.
Evaluation and clean labels set the noise floor — the Shannon bound of your data.
No model drops below it. A bigger model cannot recover information you never captured.
A governed ontology is the only thing that lowers the floor — by keeping labels clean and current.
Reviewing old labels isn't maintenance. It's the job.