SAMPLE 014
Retail04:12
Return desk handoff
Two cameras on a weekend return desk. Every override is timestamped by the associate who approved it.
Live capture / today's records
Ten thousand
decisions a day. You keep almost none of them. From here you keep every one.
Ultro Labs records the decisions your best people make, and the reason behind each one, as the work happens. You own every record. Every dataset comes with a test.
Backed by
Live decision captureProcessing
Pass 01 / 03 · Record 1042Unstructured moment
Customer messages about a wrong charge
14:08:21 · Log
“I was charged twice for the same order.”
14:08:24 · Audio
Agent pulls up the billing history
14:08:29 · Screen
Listening
Structured record
VerifiedCharge disputed
- Decision id
- UL-1042
- Human reason
- Two charges posted eleven minutes apart, same order number
- Provenance
- Log · Audio · Screen
Attribution complete1:1
The collection
Five kinds of record, taken from real work.
Every engagement produces at least one of these, tied to the person who made the call and the reason they gave.
- 01
Video
Work as it happens, framed so the decision is visible.
ExampleA weekend return desk, two angles, every override called aloud.
- 02
Audio
The reasoning out loud, marked at the moment of the call.
ExampleA rep narrating why a discount got approved.
- 03
Screen
What a person actually looked at, in the order they looked.
ExampleEight bookings moved in a scheduling tool, each with a spoken reason.
- 04
Sensor
The readings, lined up with the operator's note that explains them.
ExampleSeventy-two hours of store traffic next to the manager's log.
- 05
Log
Decisions with their inputs attached, not their outcomes.
ExampleFour hundred discount approvals and what the rep said first.
Specimen library
What a finished specimen looks like.
Real examples of finished work. Each sample says what it is, how we recorded it, and how we tested it.
SAMPLE 027
Support11:38
Support escalation call
Waveform plus a transcript marked at the moment the agent decided to escalate the case.
SAMPLE 031
Operations1440P
Annotated screen capture
A scheduler moving eight bookings in a legacy system, with a spoken reason recorded for each move.
SAMPLE 046
Retail72H
Store traffic and temperature log
Foot-traffic counts and cooler temperature from one location, aligned to the manager's note on what changed.
SAMPLE 052
RetailVERIFIED
Labeled decision log
Four hundred discount approvals with the reason the rep actually gave, not the one the system recorded.
Next
StartYour work, filed.
The specimen that matters most does not exist yet, because nobody has recorded it.
The supply problem
The web is running out of human writing.
The open web keeps growing, but the pool of attributable, human-made source material does not grow with it. Here we explain why clean records now have to be made at the source.
Open web supplySampling
Findings 03 / 03What the pool does
01 · Direction
A rising share of what gets published on the open web is written by machines, not people.
Measured by multiple independent web content studies. The direction is not in dispute.
02 · Rare tail
Models trained on model output collapse. The rare cases go first, then the specifics.
Shumailov et al., AI models collapse when trained on recursively generated data, Nature, 2024.
Source03 · Fixed pool
The pool of verifiably pre-2022 human writing does not grow. It only gets spent.
A supply constraint, not a forecast. Nothing new is being added to that pool.
Clean supply
Copy chain, generation 01 of 08
Detail kept:100%
Generation 1 of 8. 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.
What we do
We make the record that does not exist yet.
Supply line 01
Inside your walls
We record work your people already do, and the reason behind each call, at the moment they make it.
Counter camera · Spoken reason · Exception logged
A counter camera catching how an associate actually handles a return, not the order the manual lists.
1:1
Every decision, with its reason attached
Supply line 02
Across the market
The same protocol run at several sites, so a finding holds up outside one building.
Shared intake · Four locations · Variance compared
Four locations running one intake script, so a pattern is a pattern and not a local habit.
4
Locations running one script
Supply line 03
Out in the field
Our collectors go where the work is: a cab, a route, a dock. Camera and mic on from the first minute.
On route · Audio + video · Custody logged
A collector riding a service route for a week, recording the call and the reason given.
5
All five kinds of record
“A dataset without a test is a claim. A dataset with a test is an instrument.”
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.
Exclusive rightsProvenance recordEval included
Balance sheetCompounding
Year 01 / 05Data you own
AccruingUsed again with every model you try, and it stays on the balance sheet.
A model license
ExpiringRented for a term. When it lapses you are back where you started.
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 output, 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.
Warehouse floor, second shift
Copies of copies fade. Originals do not.
Start here
Tell us what to record: the work, the places, and what a good example looks like.
Every record opens with who filed it.
Backing
Backed by Y Combinator and General Catalyst.
We go inside companies and record how the work really gets done. The company owns every record.
Backing / on record02 / 02