fAIver hardware check

If a fAIver listing says "8× H100", that used to be a seller typing text into a box. Nobody checked it. This is the check.

You are being asked to run code on your own machine, so all of it is here to read. A promise that it is harmless is worth nothing. A hundred lines you can read in five minutes is worth something. Read check.py before you run it.

Running it on Windows, without knowing any of this

The instructions further down assume Linux and Docker, which is what a machine being rented out normally runs. On a Windows desktop with a graphics card in it, do this instead. Three steps, about fifteen minutes, most of it downloading.

1. Install Python. Get it from python.org, and on the first screen tick "Add python.exe to PATH" before clicking Install. That tickbox is the one everybody misses and everything else depends on it.

2. Open PowerShell — press Start, type powershell, press Enter — and paste these two lines, one at a time. The first is a ~2.5 GB download, so it takes a while.

   pip install torch --index-url https://download.pytorch.org/whl/cu121
   curl.exe -fsSLO https://raw.githubusercontent.com/jjakemaness/faiver-verify/main/check.py

3. Run it.

   python check.py

It prints a block of text and stops. It sends nothing anywhere — sending needs --report, which is not in any command above.

Two things that trip people up: it is python, not python3, on Windows; and it is curl.exe, not curl, because plain curl in PowerShell is a different command that will not do what you want.

Running it

One line, on the machine being rented:

curl -fsSLO https://raw.githubusercontent.com/jjakemaness/faiver-verify/main/check.py
python3 check.py

That prints what it found and sends nothing. Look at it first. To send it, add the code fAIver gave you:

python3 check.py --token YOUR-TOKEN --report

If PyTorch is not on the machine, add --install-if-missing and it will put it there (about 2.5 GB) rather than stopping. That flag is opt-in on purpose — nothing here decides on its own to install 2.5 GB on someone's server.

It takes about two minutes.

Deliberately not curl … | sh. That would be one keystroke shorter and would mean nothing ever lands where you could read it. It would also make one website a way to run code on every machine that ever runs this. The file goes on disk, it is the same file for everybody, and you can read it before or after.

There is a Dockerfile here too, but no image is published — build it yourself with docker build -t faiver-verify . if you would rather run it that way.

What it looks at

• What your GPU calls itself, and how much memory it reports
• How much of that memory can actually be allocated
• How long one large matrix multiplication takes
• How fast a large block of memory copies

What it does not look at

Your files. It never opens one. Not your hostname, username, IP address, serial numbers, MAC address, or what else is running on the machine.

There is no code in check.py that reads any of those. You can confirm that yourself in less time than it takes to read this paragraph — that is the guarantee, not this sentence.

What leaves your machine

Only the block it prints, and only with --report. It looks like this:

{
  "version": 1,
  "token": "...",
  "checked_at": "2026-08-09T01:17:06Z",
  "platform": "Linux",
  "driver": { "available": true, "gpus": [{ "name": "NVIDIA H100 80GB HBM3", ... }] },
  "measured": { "tflops": 780.4, "allocated_vram_gb": 79.2, ... }
}

What it leaves behind

Nothing. It runs, prints, and exits. Nothing is installed and nothing keeps running. Delete the image afterwards if you like.

What this proves, and what it does not

It proves that a machine matching the description finished a real calculation in a time only that class of hardware can manage. You can claim anything in a text box. You cannot claim to have finished in three seconds when your hardware needs five minutes.

It does not prove the machine is yours, that it is the one a buyer will receive, or that it will still be there next week. Someone could rent good hardware for an hour, pass the check, and then rent out something worse.

That is why the badge carries a date and expires. "Checked 3 days ago" is a fact a buyer can weigh. "Verified" would be a claim this cannot support, so it does not make it.

And one more, in the interest of not overselling it: because this code is public, a determined person could write their own version that reports numbers it never measured. Publishing the source makes that easier, not harder. That is a deliberate trade — the privacy guarantee is worth more than the extra obscurity, which was never much of a defence anyway. What this raises is the cost of lying: from typing a number into a box, to writing code to forge a benchmark, on an account that is also holding money in escrow with a real refund path behind it.

It is a filter, not a proof, and it is described that way everywhere it appears.