Most FPS boosters never measure anything so I built one that has to

Every gaming optimizer I tried looked the same. A big green button, a pile of registry tweaks, half of Windows turned off and a screenshot of a bigger FPS number taken on someone else's PC. Not one of them measured my own games before and after.
So I built DrDXT, a free and open source optimizer for Windows where a setting has to prove itself on your matches or it goes away.
Why one before and after proves nothing
FPS in a real match is noisy. The map, the lobby, a Discord call, the shader cache, all of that moves the number. Two games with the exact same settings can be 10 percent apart. That's how myths like disabling Nagle have survived for 15 years, nobody measures them properly.
How DrDXT judges a setting
DrDXT records your matches in the background with PresentMon. Average FPS, the slowest 1 percent, what the bottleneck was and which programs were open next to the game. It finds the game by itself and leaves launchers and Discord aside.
To test a setting it alternates games in blocks, with then without then without then with. That order cancels slow drifts like a PC getting hotter during a session. It decides after at least 3 games with and 6 without.
A gain only shows up when the two series are clearly apart, and it shows the smallest gap, not the best one. No measurable gain is a valid answer and honestly it's the most common one.
What the build refuses
I write DrDXT alone with a lot of help from AI coding assistants. I don't trust code because it looks right, so the build refuses things instead.
Every setting reads the original value before writing, and a bench applies, reads back, reverts and reads again every single one. The elevated helper can only write to an allowlist and CI recounts every write site against a document, a new write without its line breaks the build. The usual myths like Nagle, NetworkThrottlingIndex, 8000 Hz polling or realtime priority are blocked by a test. A value that couldn't be read is never reported as fine. Even the test count in the README is checked against the code.
Where it stands
It's version 0.9, C# on .NET 10 with WPF, GPL 3. It runs on my Intel and NVIDIA machine every day. The AMD code is written but stays read only until someone with a Radeon or Ryzen machine sends a report, and that's where outside help matters most. There are good first issues on the repo and the first outside contributor just landed AMD readings.
The code is in French because I'm French and I read every line, but the README and the contributor docs are in English and issues in English are welcome.
If you know Windows internals or PresentMon, or you just have an AMD PC and ten minutes, I'd really like your eyes on it.

https://github.com/ablanchard-dev/drdxt