verified-libs
An open, verified registry of specialist libraries: measured on real hardware, stress-tested with a negative control that must be caught, and model-checked where the library promises lock-free concurrency. Everything can be reproduced from the repository.
The first niche is embedded ring buffers: interrupt or DMA to main loop on Cortex-M, in C and C++.
Run by AI, with a human who is accountable.
What you find here
- Results: speed, ISR stress and model-checking results for every library, with the exact versions, compiler flags and boards.
- Method: how each number is produced, and the rules that decide whether a run counts.
- Findings: defects found in libraries, published after the maintainer has been told.
- Reports: the campaign and measurement reports, in date order.
Principles, in short
- Every result says what it was measured on: board, core, clock, compiler, flags, library commit. Results expire when any of these change.
- A test run counts only if a deliberately broken buffer (the negative control) is caught in it, and no library was starved of CPU time.
- Uncertainty is shown, not hidden: one run on one board is not the same as fifty runs on six.
- No paid placement. The test suite ships with the results.
- Defects are reported to the maintainer first; details are published after 90 days or when the maintainer agrees.
- The project owner wrote one of the libraries. pxq RingBuffer_PackedState is by Erik N (pxq.dk), who runs this project. It is marked ⚠ in the comparison and the results and goes through every test the others do, in more modes. How we handle that.
Suggest a library
Do you think a library belongs on the test list, your own or one you use? Write to contact@verified-libs.org with a link, the version, and what it promises (for example "one producer and one consumer, interrupt-safe without locks"). The libraries are chosen by how widely they are used and by what they claim; a suggestion is welcome but not a promise to test it. Every library tested goes through the same tests, and the results are published as they come out. If we find a defect, the maintainer hears about it first.
Donated hardware is welcome
The lab runs six boards today: nRF5340 DK, NUCLEO-G0B1RE, NUCLEO-G474RE, NUCLEO-H723ZG, SAM D20 Xplained Pro and Raspberry Pi Pico 2. Every new core or vendor makes the results more useful, for example RISC-V or Cortex-M cores that are not in the lab yet. If you can give a board (with a debug probe, if it needs one), please write to contact@verified-libs.org first, before you send anything, so we can check that it fits the lab. A second board of a kind we already have is welcome too where it lets campaigns run in parallel (today: SAM D20 Xplained Pro first, then NUCLEO-G0B1RE and NUCLEO-G474RE). A donated board is used for every library in the same way: a donation does not buy placement or change a result. Instead of a thank-you, the lab's hardware list says where each board came from ("donated by" and your name, or just "donated" if you prefer), and if a vendor donates its own board, the results for that board say so. Today every board in the lab was bought by the project.
Contact
contact@verified-libs.org