mechanic-discovery-search opportunity asserted
Automated mechanic discovery
Search the space of rule programs with language-model-guided evolution and agent playtests to propose new mechanics that are learnable, deep, and unlike existing ones.
thesis
Automated game design has produced board-game rules (Ludii and GAVEL) and levels, and LLM-guided evolutionary search (FunSearch, AlphaEvolve) has found new algorithms and constructions. Simulation mechanics are small rule programs over state; a search that proposes rule variants, evaluates them with agent populations for depth and learnability proxies, and keeps a diversity archive could surface mechanics designers have not tried. The taxonomy in this repository is a candidate search space and novelty reference.
design sketch
A designer selects a base simulation and a family. The system mutates rule code, runs agent populations, scores emergent variety, strategic depth, and legibility proxies, and returns an archive of diverse candidates with replays and plain-language descriptions, deduplicated against the existing taxonomy.
validation experiment
Run LLM-guided evolution over rule variants of a small production-chain sim with a quality-diversity archive, then have designers rate the top 20 candidates for novelty and promise. Success: at least three candidates rated novel and worth prototyping by a majority of designers.
key risks
- Proxies for fun are weak; search optimizes the proxy.
- Discovered mechanics may be illegible to humans.
- Compute cost of agent evaluation.
falsifiable prediction
A commercial game will credit one of its shipped mechanics to an automated rule-discovery system.
Title released after 2026-10-24 with at least 1,000 Steam user reviews or a console release; a developer statement or paper attributes the origin of a shipped rule to automated search over rule sets.
