Made, not found.

Your company made ten thousand decisions yesterday.Almost none of it survives.

Ultro Labs runs experiments inside your company to capture the record before it fades: clean, human, multimodal data on how the work actually gets done.

Start an experiment.
A clean digital signal degrading into scattered noise, with one orange node marking the lock

The supply problem

The human signal is getting harder to isolate.

The open web keeps growing, but the pool of attributable, human-made source material does not grow with it. Three facts explain why clean records now have to be made at the source.

Continue / three signals

A rising share of what gets published on the open web is written by machines, not people.1

1. Measured by multiple independent web content studies. The direction is not in dispute.

Models trained on model output collapse. The rare cases go first, then the specifics.2

2. Shumailov et al., AI models collapse when trained on recursively generated data, Nature, 2024. Source

The pool of verifiably pre 2022 human writing does not grow. It only gets spent.3

3. A supply constraint, not a forecast. Nothing new is being added to that pool.

The decay

One paragraph of real operational knowledge, copied eight times.

Generation 1 of 8

Fidelity
100%
12345678

Generation 1. Marisol has run the returns desk for eleven years. She can tell from the tape on a box whether the customer repacked it at home or in the parking lot, and she knows the difference matters: parking lot returns are impulse regret, home returns are real defects. None of this is in the manual.

Quality reviewers verifying field cameras, record sheets, and captured footage in an operational data lab

From noise to evidence

Clean is not a look. It is a chain of evidence.

Every record is captured at the source, tied to the person and context behind it, then tested against a measurable outcome.

  • Human origin
  • Full context
  • Built-in eval

What we do

We make the record that does not exist yet.

We instrument the work your own people already do, and record the reason behind each call.

A floor camera catching how a technician actually clears a jam, not the order the manual lists.

Overhead capture of a technician clearing a machine while ceiling sensors trace the actual route

The same protocol run at several sites, so a finding holds up outside one building.

Four depots running one intake script, so a pattern is a pattern and not a local habit.

Four logistics depots connected to one shared capture protocol

Trained collectors capture work where it happens. Multimodal from the first minute.

A collector riding a service route for a week, recording the call and the reason given for it.

A field collector recording a service technician beside a mobile capture vehicle

Lab layer

Every dataset ships with its own eval.

Held-out cases turn a collection into a testable instrument, measured against your own ground truth.

Specimen library

Five sample types, taken from real work.

The asset

Data is capex now.

A model license expires. A dataset does not. Data collected on your own ground stays on the balance sheet, it is used again with every model you try, and it gets more valuable as the open supply gets worse. You are not renting an answer. You are buying an asset, and you own it outright.

YEAR 1YEAR 3STILL YOURS

Evals

If it is not measured on your own ground truth, it is a guess.

From the manifesto

The open web is filling with text that no person wrote and no person checked. It is cheap to make and it reads well enough to pass. Every month there is more of it, and every month the average page is a little further from anyone who did the work it describes.

Models trained on that output drift. When a system learns from its own reflections, the rare cases go first, then the specifics, and what is left is a confident average. The finding has a name and a citation: Shumailov and colleagues published it in Nature in 2024. The practical version is simpler. Copies of copies fade.

Read the manifesto

Tell us what your company knows that no dataset contains.

hello@ultrolabs.com