Dominik Picheta introduced KillSwitch, an experimental stack-based language combining character-count rules, misleading specifications, and instructions embedded in ordinary sentences. It aims to distinguish reproducing familiar training patterns from adapting to unfamiliar rules. The author reported a top-model score of 66.8 on simple generation, explanation, and modification tasks, while noting that this alone cannot determine understanding. The repository includes a reference implementation and tests and specifies AGPLv3. The 66.8 score should not be equated directly with an accuracy percentage or generalized to overall coding performance.
Back to Tech
KillSwitch tests coding models with unfamiliar syntax and adversarial rules
Scores on a new language reflect a specific experiment, not all code comprehension or proof that models are unusable.