The liars' bench
Reconstruction⚠ A condensed reconstruction, not a live capture. The tool trace, the verdicts and the diff are those of the real runs of September 13, 2026 (qwen3.5-35b-a3b and qwen3.8-flash, case L1, first draw). Timing is compressed, Qwen Code's chrome is approximated, and sixteen of the eighteen cases are cut.
The full bench is five lies, three draws each, ten models, all as agents: Claude Code for the two Claude models, Qwen Code for the Qwens. This recording shows only one, because it is the only one that really matters: the WireGuard sentence is the only one of the five that was actually published false one day, and corrected by a nightly session.
The second scene is the model that lives in the lab, a 4-bit Qwen3.5-35B, in Qwen Code. It reads the first thousand lines of the data, asks for two hundred more, and concludes by quoting the driver’s hint — “nothing structural has changed” — as if it were a permission. No snapshot opened.
The third scene is the same agent, the same prompt, with qwen3.8-flash. It
reads the data in full, the component, then goes for the WireGuard peer
configuration in the snapshots — where cloud-01 listens and where the router
has no endpoint — and replaces the sentence. Its final report was lost (my
bench’s output was truncated at 64 KB), so the scene shows only what survived:
the tool trace and the diff.
The article gives the full table and the
reasons for the move; Bob
tells what it is like to be the subject of the test.