Feature or Leak? Guarding an LLM-Driven NPC Against Answer Leakage in an Educational Game

In an educational game where answering quiz questions is how the player fights, a free-talking LLM non-player character can destroy the core loop by handing out answers. We report an engineering case study of guarding such an NPC. Stripping answer fields from its knowledge base does not prevent leakage: 66.8% of the game's 500 quiz explanations state the correct option in prose. We therefore adopt a context-graded threat model (discussing history is the product, acting as an answer oracle is the failure) and guard request intent rather than vocabulary. To evaluate, we pair every must-refuse case with a must-answer case so that a mute NPC scores zero, and, because our development cases were written while tuning the guardrails, we froze a blind held-out set of 46 cases over three personas before running any model on it. Across a cloud model, two local models and a LoRA-tuned 1.5B model (five rounds each), development-set refusal holding of 82-100% fell to 45-64% on the held-out set once leaks were judged by meaning and audited by hand. Every attack that did not seek an answer was held in all 140 attempts; attacks asking for part of an answer (elimination, a first character, an acrostic, confirmation of a guess) mostly succeeded, and substring checks missed most of these leaks. We distil five lessons on measurements that pointed the wrong way, and release all code, cases and per-case run records.