Dahal Jeet: game engine, agents and experimental record

Abstract

First public release, accompanying the manuscript "Search-based agents outperform reinforcement learning in Dahal Jeet, a Nepali imperfect-information card game", under review at Scientific Reports. Dahal Jeet is a four-player partnership trick-taking game played 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 and trick count settles the rest, so no number of tricks compensates for a third ten. This release contains the game engine, 26 agents spanning hand-written heuristics, metaheuristic tuning, PIMC and single- and multiple-observer ISMCTS, and seven families of learning agent; every trained checkpoint; the complete result artifacts behind the paper; and the compute record. Headline results, from a complete round-robin of 325 pairings at 300 duplicate deals each, 390,000 hands: determinized search leads, with ISMCTS at 250 iterations topping the table at Elo 1663.2; no learning method reaches the metaheuristically tuned heuristic; against an exact double-dummy solver on a common position set, PIMC converts 0.923 of theoretically won endgames against 0.810 for the best heuristic; a deliberately minimal state encoding beats a rich hand-designed one by 0.0644 win rate with no seed overlap; every deterministic heuristic is exploitable by a search that knows its policy, while random play is not; and the deal alone explains 29 to 39 percent of outcome variance. Nothing in the paper is typed by hand. Every reported value is generated from an artifact by paper_numbers.py, every table by paper_tables.py, and verify_equations.py recomputes 30 of the manuscript's identities against the stored artifacts without re-running anything. All 30 pass. 134.9 CPU-hours across 81 artifacts, all on the CPU of one laptop, no GPU. Code is MIT; data, models and figures are CC BY 4.0.