Interactive code evaluation

One green test can lie.

A generated kernel passes the fixed test. Choose the next input that exposes what it missed. The browser runs both the candidate and a reference for every choice.

Five original teaching examples. About three minutes. No model call or upload.
Case 1 of 5Caught: 0
Link to this case
01 / 05

Fixed test passed
Candidate sketch

Why this matters

Correctness needs a wider input regime.

These browser-run sketches use original examples, not GPU kernels. They illustrate the fixed-shape testing problem studied in Dipankar Sarkar's work on LLM-generated GPU kernels. His published corpus pairs deliberately faulty variants with correct controls and uses seeded tests across shapes and dtypes.