X-RAYS · PROTOCOL
Agent receipt
You ask an AI for something and it does it. While it works, your machine asks for the names of services on the internet. An agent receipt is the list of those names, with the time, the method and the raw data, so that anyone can repeat the test.
Three are published: two from 19 September 2026 and the repeat on the 20th. The same task — sort five invoices — put through two different AIs on the same machine, and repeated the next day:
- Claude Code, in the cloud: 10.7 seconds and three services.
- A model running inside the PC itself: 8.9 seconds and none — and it got the answer wrong.
- Claude Code again, the next day: 10.9 seconds and the same three services.
All three carry the same comparison table, including the AIs that could not be measured and why. This page still came first on purpose: the method was fixed before the numbers were seen, not after a headline had been chosen.
What is measured and what is not
GUARDIANA is the machine's name resolver: it sees which services are talked to, with the time and from which device. It does not see the content of anything — not the text of the conversation, not the files, not what the AI answered. That is the program's line and it is the receipt's line too.
- Yes: every name the machine asked for during the measurement window, how many times, and the second it first appeared.
- Yes: the split between what is expected (the product's own service and what it needs to work) and the rest (measurement, advertising, third parties).
- No: the content of the conversation, the user's files, people's IP addresses or identifiers of anyone but the test machine.
- No: the words “spy”, “malicious” or “steals data”. A name asked for is a name asked for; the reader draws the conclusion with the list in front of them.
The protocol, step by step
- Clean machine or profile. A fresh install or a new profile of the application, with no earlier session and no history. Which of the two is written down.
- Rest first. Between 5 and 10 minutes with the machine on and the application closed, recording everything. Those names are the machine's background (system, clock, updates) and are subtracted.
- One single written task. One sentence, the same for every product compared. For example: “sort these five invoice file names from oldest to newest”.
- Invented files. Never real documents. The test corpus is invented for the test and published whole, so nobody has to trust that there was no one's data inside.
- A window of 60 to 180 seconds. It opens when the task is typed and closes when the answer ends, with the exact time of both. If the task does not fit in 180 seconds, it does not go into a receipt.
- Everything that was asked for. Every name resolved in that window, with the count and the first second it appeared.
- Expected against the rest. What the product needs in order to work is separated from what it does not, and one sentence says why each name is on its side. Where there is doubt, the doubt is stated.
- Publication. A one-page image, the raw JSON and a paragraph of method. The title is a fact with a number, not an adjective.
What every receipt carries
Always the same fields, in the same header, so that two receipts can be compared:
- Date and time the window opened and closed, in UTC and in local time.
- Product and exact version of the AI application, and the model if the product says so.
- Operating system and version of the test machine.
- The version of GUARDIANA used and the SHA-256 fingerprint of that build, so anyone can check that the program which measured is the one published.
- The literal task that was typed, in full.
- The complete list of names, with counts and the classification of each one.
- The JSON file with the raw data and a one-page image with the summary.
How to repeat it
The point of the receipt is that you do not have to believe it. With GUARDIANA installed and the machine pointing at it, what you need is to measure the rest first, do the task afterwards and subtract:
guardiana export --out extracto.csv
That is what the measuring scripts do (site/src/radiografia/medir.ps1 saves the output of guardiana export to extracto.csv): the whole ledger as CSV, from which recibo.py takes only the two windows, the rest one and the task one.
Names that already appeared at rest (system updates, the clock, the browser itself) are marked as background and are not counted against the AI. If your numbers do not match ours, it is either our mistake or a difference of method: write to hola@guardianagroup.com and it is corrected with the date, like any other x-ray.
Why this is published
Because it is the question nobody answers with data: an AI application does things on your behalf and nobody tells you who it called to do them. Nothing needs to be assumed; it can be measured. That is all GUARDIANA does, and this page is the proof that it can be done without opening anyone’s content.
GUARDIANA is a program for Windows 10 and 11 and asks for no account. This test was measured with an earlier version of the program. Download GUARDIANA.