Quality control for what AI learns from
Break the loop before it breaks your model.
AI learns from the internet. More and more of the internet is written by AI. We check that what your model learns from is still human, still clean and still whole.
01Why this exists
AI is starting to learn from itself. That's a problem.
Every AI model is built by reading an enormous amount of human work: books, articles, code, conversations. Now a growing share of what's online is written by AI. So the next model learns partly from the last one. It's a copy of a copy, and copies lose detail.
Human originalMy grandmother's rasam, soured with tamarind from the Madurai market
Copy 1Grandmother's rasam with tamarind
Copy 2A traditional tamarind soup
Copy 3A delicious soup
Human originalHaar: what east-coast Scots call the cold sea fog that rolls in off the North Sea
Copy 1Haar is a Scottish word for sea fog
Copy 2A Scottish word for fog
Copy 3Fog
Human originalHonestly? The sequel is worse, and I will die on this hill.
Copy 1The sequel is not as good as the original.
Copy 2The sequel received mixed reviews.
Copy 3Many people enjoyed the film.
Nothing in the last column is false. That's the trouble. It's just emptier, and the emptiness is hard to measure. A model trained this way doesn't break. It slowly forgets what made it useful.
- 01
Copies lose the details first.
The rare fact, the small language, the distinctive voice. A model trained on copies still sounds fluent. It's just blander, and more often wrong about the things few people know.
- 02
It's hard to see from the inside.
Standard tests check averages, and averages hold up. By the time anyone notices what went missing, the model has already shipped.
- 03
No one checks from the outside.
Labs grade their own models. Data vendors grade their own data. There is no independent inspector for what AI learns from. Yet.
The technical name
Researchers call this model collapse. It isn't a crash. It's a slow drift toward the beige middle that standard evaluations, built around averages, barely register.
What we do, in one sentence
Ouroboros is the independent inspector for what AI learns from. We test it, certify it and publish the scores.
Think of a food-safety lab for AI's diet. Your model is what it eats. We make sure the people building it, buying it and regulating it can see exactly what that is.
Why the name
“The serpent that eats its own tail.”
It is the oldest picture of a system feeding on its own output, which is how model collapse begins. In the old books it also stands for renewal. We work on the second meaning: a loop that keeps taking in fresh human signal instead of consuming itself.
02What we do
Test it. Certify it. Score it.
Ouroboros Antivenom
Adversarial data red-teaming
We try to break your data pipeline before anyone else can, safely and inside your own systems.
Antivenom is made by controlled exposure. So is a resilient pipeline. We test whether your ingestion, filtering, RLHF and evaluation stack holds against flooding and synthetic saturation, inside your environment and under your rules of engagement.
- Findings report
- Remediation plan
- Retest
Ouroboros Seal
Human-data provenance certification
We certify that data is genuinely human-made, so the people who made it can prove it and charge for it.
The Unbroken Ring: an audited chain of custody attesting that a dataset or archive is human-made, consented and uncontaminated. A mark publishers can carry into licensing, and buyers can check before they train.
- Dataset audit
- Certification mark
- Annual renewal
OuroBench
Public slop-resistance benchmark
A public scoreboard for how well each AI model remembers the rare stuff.
The tails go first: rare facts, dialects, craft knowledge, a specific voice. OuroBench measures whether a model keeps them across versions. The public leaderboard is free. A private suite runs before you ship.
- Public leaderboard
- Private eval suite
- Version-over-version diffs
Molt · Continuous contamination monitoring that sheds bad data at intake.
The Coil · A public registry where seals, datasets and OuroBench scores live.
03How it works
An engagement is a loop that closes.
- Step 01
Scope
We agree in writing on systems, data and rules of engagement. Nothing runs outside that line.
- Step 02
Contained testing
Testing runs inside your environment, on your infrastructure. Adversarial material is created there and destroyed there.
- Step 03
Findings
A plain report: what held, what gave way, what it would cost you and what to fix first.
- Step 04
Retest
After you remediate, we run it again. The loop closes when the fix holds.
04Who it's for
For anyone whose model is only as good as what it ate.
Model builders
Labs that want an outside read on their data before a release, and evidence to hand enterprise buyers.
Enterprises
Teams fine-tuning or running retrieval over internal corpora, where quiet contamination turns into a regulated output.
Governments
Safety institutes and procurement offices that need a neutral, repeatable evaluation.
Publishers
Archives and rights-holders who want to show their work is human-made, and price it that way.
05Why now
A shift happening under the surface.
The human supply is finite.
Researchers at Epoch AI have projected that AI training could use up the stock of quality public human text within this decade. What's left gets more valuable, and more worth protecting.
Synthetic content is now the cheap default.
Writing a million words used to take a million hours. Now it takes an afternoon. Every crawl, search index and company knowledge base is absorbing the difference.
The rules are arriving.
Europe's AI Act asks makers of general-purpose models to publish summaries of their training data. Summaries invite a question: who checked?
Human work is becoming an asset.
Publishers and creators are starting to license their work to AI companies. Proof that something was made by people makes it worth more, and puts value back with the people who made it.
The pattern
Every new layer of technology has eventually needed an independent checker. Auditors for company accounts. Underwriters Laboratories for electrical products. SOC 2 for cloud software. Training data is next.
06OuroBench
A public score for what models keep.
How much of the distribution's tail a model keeps after training on a contaminated web. Higher is better. Free to read, open in method, governed at arm's length.
| # | Model | Score | Rare facts | Dialects | Style range | Δ prior |
|---|---|---|---|---|---|---|
| 01 | Model A | 0.81 | 0.84 | 0.79 | 0.81 | +0.02 |
| 02 | Model B | 0.77 | 0.80 | 0.74 | 0.78 | +0.01 |
| 03 | Model C | 0.74 | 0.77 | 0.70 | 0.75 | −0.01 |
| 04 | Model D | 0.70 | 0.71 | 0.66 | 0.72 | +0.03 |
| 05 | Model E collapse flag | 0.60 | 0.66 | 0.52 | 0.61 | −0.12 |
| 06 | Model F | 0.58 | 0.60 | 0.55 | 0.58 | ±0.00 |
| 07 | Model G | 0.50 | 0.54 | 0.47 | 0.50 | −0.02 |
Demo data · Model names and scores are invented placeholders to show the format. No real model has been measured.
07Where this goes
From a service to a standard.
Stage 1
Test
Red-team engagements with model builders. Paid work that builds the know-how, the data and the first reference customers.
Stage 2
Certify
The Seal becomes the mark buyers look for, like SOC 2 for cloud software. Publishers certify their archives to license them at a premium.
Stage 3
Register
The Coil: a public registry of certified datasets and scores. More certified data brings more buyers, and more buyers bring more data.
Stage 4
Reference
Regulators, insurers and procurement teams cite the standard. That's the point where an auditor becomes infrastructure.
Plan, not a promise · Stages and sequencing are illustrative
08Principles
An auditor is only worth its independence.
Neutral
No lab owns us.
We won't take investment or exclusivity that would let one customer grade another. The public benchmark is governed at arm's length.
Contained
Adversarial data never leaves the client.
We publish measurements, never methods. Test material stays inside your environment and is destroyed at the end of an engagement.
Public benefit
A Delaware public benefit corporation.
The purpose is written into the charter: the integrity and diversity of the human information commons.
09Why us
Built by people who measure what's rare.
Arvind Nagarajan
Founder & CEO
Arvind's work has centered on one question: how do you find and measure what's rare? He built an adaptive assessment designed to spot exceptional students that standard tests miss, which is work that lives in the tails, exactly where AI collapse hits first.
He's also a co-founder and operator in education, building institutions that grow human potential: the original source of every dataset these models depend on.
“I started by studying how AI could be made worse. I came out convinced the defense matters more. Someone has to protect the human signal these models live on.”
Founding CTO
HiringA technical co-founder with published work in data curation, robustness or model collapse, or who has led data integrity inside a major lab.
An auditor is only as credible as its technical name. This seat comes first.
Start a conversation10Contact
Request an audit.
Tell us what you train on and what you ship. We'll reply with a proposed scope, not a pitch deck.