REAL UPSTREAM BUGS · NVIDIA NEMOTRON ON NEBIUS TOKEN FACTORY · REAL BROWSER VERDICT
Pick a bug maintainers fixed by hand.
Watch Nemotron fix it blind.
Each bug was reported in a real open-source library. For bugs the maintainers already fixed, Tabstitch pins the code from just before the fix and freezes the user's keyboard task; the maintainers' own fix must pass that task first, so the test is fair, and Nemotron gets only the buggy source. Bugs still unfixed upstream are reproduced on the library's current code. Nemotron 3 Super tries first; if the browser rejects its patch, Nemotron 3 Ultra gets the exact failure. Chromium replays every path and gives the verdict, not the model.
Latest live run
LIVE RUN
The bug, reproduced
Compare with the maintainers' own fix
This is the published upstream change. It was used only to check the task before the run, and it was never sent to the model.
Loading…
Full evidence: every capture, the raw model request and response, and a proof archive →
Still open upstream
Reproduced on each library's current code. No maintainer has fixed these yet.
Real bugs
Scoreboard
One blind run per bug, routed Super → Ultra, recorded before this page was published.
| Library | Bug | Maintainers' fix passes the task | Nemotron, blind | Passed on | Evidence |
|---|
Verified only for the frozen tasks and recorded conditions. The page around each library is author-written; the library code is the unmodified upstream source from before the fix. Public fixes may appear in model training data, so a pass shows the verify loop working on real code. It does not prove the model never saw the fix. Each run's raw request is checked for lines unique to the maintainers' fix.