Rundale Bench

First-person NPC dialogue in the voice of Brigid O'Brien — a 42-year-old midwife in 1820 rural Ireland. Probes period-accurate vocabulary, in-character voice, en-IE / ga-IE code-switching, and refusal of non-Latin script. The headline quality slice — drives the main leaderboard.

Short in-character one-liners NPCs emit in response to nearby game events (weather shifts, arrivals, gossip). Tests whether the model can stay in voice under tight token budgets without drifting into narration.

Structured world-tick outputs. The model emits JSON describing NPC state updates (mood, goal, current action) given a scene. Tests schema compliance and plausible micro-simulation — the engine runs hundreds of these per game day.

Deeper structured sim: multi-step NPC plans, conditional triggers, longer JSON. Same schema-validation + plausibility bar as tier2, but the model must compose several intents coherently.

Irish-language (Gaeilge) fluency. Eleven prompts in Irish probe natural syntax, idiom, grammar, task-fulfilment, and resistance to falling back to English. Decoupled from the dialogue slice so models that fake en-IE can't fake ga-IE.

Deterministic player-input parser. Maps natural-language input ("go to the pub", "tell Mary I saw her cow") to {intent: move|talk|look|interact|examine|unknown, target, dialogue}. Exact-match graded — no LLM judge, no axes; the only slice driven entirely by deterministic scoring.