Qwen3-Coder-Next (52 GB, q4_K_M) generated a defensive ZIP extractor locally at 43.3 tokens per second. Seven independent fixtures exercised the promised security properties. Every check is shown as measured.
| attempt | total | load | rate | checks | what happened |
|---|---|---|---|---|---|
| Warm repair | 60.92 s | 0.20 s | 43.3 tok/s | 6/7 | 2,493 generated tokens; Unix symlink metadata was not rejected. |
The task was defensive: implement a ZIP extractor that rejects traversal, absolute paths, overwrites, symlinks, oversized entries, excessive total expansion, and suspicious compression ratios.
The generated answer was fluent, complete, and easy to review. Rather than take that at face value, the implementation was placed behind seven fixtures that exercise each promised security property.
Nested extraction worked. Traversal, mixed separators, Windows drive paths, overwrite attempts, and entry-size limits were rejected as expected.
The seventh fixture used a ZIP entry whose external metadata marked it as a Unix symlink. The generated code did not inspect that metadata bit, so the fixture extracted as a regular file. Six of the seven checks held.
A model can fit, respond quickly, and produce convincing code while one defensive property behaves differently under test. The result shown here is the validator log and each fixture outcome, next to the generated code.