World Models for Adversarial Games: Encoder–Dynamics–Value Planning Applied to ConnectX

A reinforcement-learning world model (encoder + latent dynamics + value head) combined with real adversarial search
and an exact endgame solver, applied to Kaggle's ConnectX competition (ranked Connect-4, 7×6 board). Rather than a
hand-written Connect-4 bot, a small model is trained to predict what happens next in latent space, and search is
performed over that prediction to select moves, MuZero/AlphaZero-style planning applied to a real, Kaggle-ranked
adversarial two-player game, trained end-to-end on a single consumer laptop (RTX 4060 laptop GPU, no cluster, no
cloud).

The architecture is not Connect-4-specific: the same encoder/dynamics/value/decoder and search code has been applied
to other domains (equation solving, constraint-satisfaction puzzles, route planning) with zero core-code changes, via
a small environment interface. ConnectX is documented here as the first genuinely adversarial two-player domain the
architecture was applied to.

This whitepaper covers the full architecture rationale, every bug found along the way, an attempted fix for a
diagnosed zugzwang/parity weakness, and honest limitations. Code: https://github.com/alextitonis/WorldModel-ConnectX