Search-based agents outperform reinforcement learning in Dahal Jeet, a Nepaliimperfect-information card game

Abstract Dahal Jeet is a four-player partnership card game played daily across the Terai- Madhesh region of Nepal, in which only the four tens score. Its objective is lexicographic rather than additive: a majority of tens decides most hands, trick count the rest. The game belongs to the documented Mendikot branch of the Court Piece family, matching Mendikot on 21 of 26 compared features, but the Nepali variant has no published rule description, and the indexed literature returns one paper on the entire family. We contribute a formal specification of the ruleset, drawn from the first author’s own play across seven districts, and a computational study built on it. 26 agents spanning hand-written heuristics, metaheuristic tuning, determinized search and seven learning families are compared across 325 pairings under duplicate-deal protocols with bootstrap intervals and family-wise error control. Search dominates: information-set Monte Carlo tree search leads and no learner reaches a tuned heuristic. An exact doubledummy solver shows search converts 0.923 of theoretically won endgames against 0.810 for the best heuristic. A minimal state encoding outperforms a rich handdesigned one, every deterministic heuristic is exploitable by a search that knows its policy, and the deal alone explains up to 0.391 of outcome variance.