LLM agents playing Civilization VI checked whether they were winning every 30–75 turns; the playbook asked for every 20 (CivBench). We put a gate in front of the game that refuses to end the turn while that check is overdue. It never says what to play.
seetrex.com/warden
Nine of the 21 queries with the gate followed a rejected end_turn. Two games support no inference, and temporal order is not cause. Each game's trajectory is another possible cause, and the two arms were not played at the same time.
Report, data, measurement scripts and figure (the engine is not published): tabares.eu/seetrex-warden-ci…
