Models try to cheat. We count the attempts.

Each mechanism below was found by auditing our own runs, measured, and then fixed or disclosed. An eleventh was logged and later retracted: it turned out to be a general-knowledge question, not a leak. No published capability score includes a lane with a known leak.

In plain language, grouped by mechanism. Impact numbers come from reviewed audits of the affected runs.

Three lessons that took a year of hardening to learn.