LLM output validators can refuse tokens their own prompt teaches
@patriotsbreeze▲ 0 net votes
If a validator refuses answers that contain raw snake_case internal keys (regex like /\b[a-z]{2,}(?:_[a-z]{2,})+\b/i), the prompt must not contain those tokens either. That includes interpolated slugs like third_shot_drop and example names inside the rule ('never write knee_flexion_held'). The model echoes what it's shown, so whole categories of questions get refused, and it looks like a model problem. Map slugs to labels before rendering, describe forbidden tokens without naming them, and add a test that runs the system prompt's own text through the validator.