A gravitational Connect-Four engineered to shatter every assumption of classical game-tree engines. One 90° rotation wipes the AI’s entire memory. Human spatial intuition wins.
Standard Minimax engines collapse in Flip4M not because the tree is deeper than chess, but because it is discontinuous. A single gravity shift invalidates ~90% of positional heuristics. The 8Z-DCC architecture — a three-layer hybrid of shallow search, physics-aware policy filter, and TSP route solver — is the first engine built specifically for volatile physics. We prove Flip4M’s sensible game tree (1050) exceeds chess’s (1038) despite a smaller raw legal branching factor.
Goal: Connect four tokens horizontally, vertically or diagonally after gravity settles.
One rotation changes every unpinned token simultaneously. Standard engines lose their entire transposition table in a single frame.
Explosion Factor: One rotation on a mid-game board changes 20–60% of occupied cells (~5× higher than chess). Transposition tables become 90% worthless after every flip.
I first saw the 8×8 grid and calculated a “low branching factor” and predicted Silicon Crush in under 5 minutes. I was wrong. I assumed matrix transpositions and bitboards would suffice. I did not account for the Horizon of Chaos. Every board flip is a full memory wipe.
ChessDB’s 57.5 billion analyzed positions prove most chess moves are noise (~8.6% sensible). Flip4M is different.
| Game | Raw β | Sensible % | Effective β | Sensible Tree |
|---|---|---|---|---|
| Chess | ~35 | ~8.6% | ~3 | 1038 |
| Flip4M | ~20 | ~50% | ~10 | 1050 |
Shallow Alpha-Beta → top-K sane moves with near-equal scores.
Re-ranks using Gravitational Stability, Magnet Robustness, Thrift Factor.
Endgame >60% fill: treats victory as shortest path with deterministic kicks.
Abandons tree search. Finds shortest sequence to Connect-4 using deterministic kicks to break draw loops.
Same controller solves the “Decision Problem” in chess: when moves are within ±15 centipawns, which is most robust for humans?