About DriftCaught
We are a subscription service that owns evaluation of your AI features: the golden dataset, the scorers, the judge calibration, the reruns on every model update, and the monthly report. We have no case studies at launch and we will not invent any.
Your AI feature can change every time the model does. Our job is to catch it before your users do.
What we do
A team ships an AI feature. Somebody writes a few test cases. The feature works. Three months later a provider ships an update, or a prompt gets edited, or the inputs users send drift away from the ones anybody tested, and nobody can say whether the feature got better or worse.
Evaluation is not a project that finishes. It is a process that has to keep running, and the reason it stops is never that the tools are bad. It stops because nobody owns it. We take the dataset, the thresholds, the judge calibration and the reruns, and we run them inside your stack with your tooling.
You can read exactly how, before you pay anything. Eight protocols, published in full, versioned, cited step by step.
Company
- Legal entity
- DriftCaught Ltd
- Form
- Cyprus private limited company
- Headquarters
- Cyprus
- Domain
- driftcaught.com
- Contact
- hello@driftcaught.com
- Reply within
- 24 hours
The rules we hold ourselves to
A provider change without the provider's own source is not published
The change log is the one asset on this site that has to be right. Every entry links to a page on the provider's own domain. We reject announcements we cannot reach there, whoever else reported them, and the build fails if an entry slips through without one.
We publish no measurement of any model
No benchmark, no leaderboard, no score. We restate what providers announced and we cite published practitioner guides. Measurements we make belong to the customer whose dataset produced them, and they stay there.
Every number has a source you can check
There are 45 items on the evidence page, each with a publisher, a URL and the date we read it. Five claims in our own specification failed verification and were dropped rather than softened. They are listed, with reasons.
The limits are published as loudly as the service
What we cannot do has its own page, is repeated on every plan, and is in the terms. A bounded service you can check is worth more than an unbounded promise you cannot.
The tools are good and we use them
Braintrust, Langfuse, Arize Phoenix, DeepEval, promptfoo, Ragas, Weights & Biases. We do not sell software and we have nothing to gain from criticising theirs. What teams are missing is not better tooling, it is somebody whose job it is to keep the dataset, the judge and the reruns current.
Change log
63
tracked provider changes, every one with the provider's own source
Evidence
45
sourced items, each with the date we read the page
Our own measurements published
0
and that is on purpose
Change-log entries restate provider announcements with sources and dates. Impact notes cite public practitioner reports. The site publishes no measurements of its own on any model.
What we are not
Not a benchmark. Not an eval tool vendor. Not a model shop. We do not train or fine-tune, and we do not review outputs for policy or legal decisions. If what you need is a tool, buy one and keep the money; several good ones are named on this site.
What we cannot do
We cannot make a model deterministic or guarantee its outputs. We measure against an agreed dataset and thresholds and report what changed. We do not train or fine-tune models, and we do not replace human review for policy or legal decisions. Coverage is bounded by the dataset and the plan.
Questions about us
Why does this service exist?
Do you have case studies?
Are you an AI company?
Who is behind it?
What would make you refuse a customer?
Found a claim on this site we cannot support? Write to us. We will correct the page or cut the claim, and the evidence page exists so you can check.