The Tetris Proof: Real-Time Neurosymbolic Superiority

Interactive Arena

Autoplay shows the reference controller. Play yourself hands you the keys; the gate then only counts the moves it would have vetoed. Click the benchmark button below to run 20 fixed seeds (1000–1019) to 40 lines on your CPU and write the results into the meta bar and Section 3.

1. The Combinatorial Trap: Where Autoregressive Generation Hits Physics

A branching factor that compounds

With hard drops from spawn and no spins, a piece has at most 34 distinct placements on an empty 10-wide board (T, J and L), 17 for I, S and Z, and 9 for O. The mean over the seven pieces is about 23. These counts come from the arena engine itself and are pinned by a unit test. Looking four pieces ahead therefore spans up to 34⁴ = 1,336,336 placement sequences, before the random pieces that follow. Offline Tetris is NP-hard even to approximate [1], so any real-time controller must prune, and must prune well.

The physical clock

At 60 Hz a frame lasts 16.7 ms. Under 20G gravity the piece reaches the stack within a single frame, so the only time left is the lock delay: a fraction of a second in guideline games, less in arcade 20G modes. A controller that writes its move as text pays for that move token by token:

This is arithmetic, not a measurement. Take a move written as about 40 output tokens and a decode rate of 50–100 tokens per second: decoding alone costs 0.4–0.8 s, before the first-token wait, the round trip and any hidden reasoning tokens, which reasoning models spend by the hundreds or thousands per step. Even under these favourable assumptions one decision spans dozens of frames. We did not time any hosted LLM for this page; replace the assumptions with your own numbers and the bound follows.

Spatial rules the decoder must re-derive every step

A text-generating agent predicts its next token from a serialized board. Nothing in that loop checks that the named placement is legal, reachable through the SRS kick table, or free of new holes: collision, rotation and line clears are exact discrete rules, and the model has to reconstruct them from text on every step. We have not measured an LLM's illegal-move or hole rate here, so we make no numeric claim about it. The engineering conclusion does not need one: whatever proposes a move, neural or symbolic, needs an exact checker downstream. The arena implements that checker.

2. The Dual-Track Neurosymbolic Architecture

Gen-Zero splits control into a fast neural proposer and an exact symbolic verifier, with a search between them that never leaves structured state. The badges say what exists today and what does not.

flowchart LR
    S["Board state + Next queue"] --> T1["Track 1: Mamba-2 reflex<br/>ROADMAP, not implemented"]
    S --> T2["Track 2: Lookahead search<br/>beam 16, depth 3, zero tokens"]
    T1 -.-> R["Ranked proposals"]
    T2 --> R
    R --> G{"Track 3: gate<br/>admit iff delta_holes <= 0"}
    G -- pass --> A["Place piece"]
    G -- "veto: next proposal" --> R
    G -- "no admissible move" --> B["Least-bad move<br/>+ breach counter (+ warning in autoplay)"]
    B --> A
    classDef roadmap fill:#12151A,stroke:#8E949A,color:#8E949A,stroke-dasharray:5 4;
    classDef live fill:#0e1218,stroke:#2FE3F0,color:#ECE7DC,stroke-width:2px;
    classDef gate fill:#0e1218,stroke:#FFA51F,color:#ECE7DC,stroke-width:2px;
    class T1 roadmap;
    class S,T2,R,A live;
    class G,B gate;

3. Empirical Benchmark Scorecard

The Gen-Zero column fills in when the benchmark in the arena finishes. The other columns are empty on purpose.

Why the LLM columns are empty. We have not run these agents under a common harness, so we publish no numbers for them. Filling the cells from vendor pages or intuition would be the kind of claim this page exists to refuse. A harness that drives each agent through the same 20 seeds and the same gate is future work.

Method. Seeds 1000–1019, seeded 7-bag, stop at 40 lines or 1,000 pieces. Search depth 3, beam 16, then the Δholes ≤ 0 gate. Decision time is performance.now() around search plus gate. Browsers coarsen that clock (the arena shows the resolution it measured), so p50 and p99 may be quantized; the batch mean (total time ÷ decisions) is not. Throughput here is compute-only: decisions per second of CPU time, not the animated play speed. Results depend on your device.

4. From Blocks to the Industrial World

What carries over is the pattern: a fast proposer, a search that stays in structured state, and an exact gate that can say no. Gen-Zero is not deployed in any of the domains below. Each card names the gate the domain would need and what this page does not show.

References

1. [1] E. D. Demaine, S. Hohenberger, D. Liben-Nowell. Tetris is Hard, Even to Approximate. COCOON 2003.
2. [2] I. El-Ashi. El-Tetris: An Improvement on Pierre Dellacherie's Algorithm. 2011.
3. [3] T. Dao, A. Gu. Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality. ICML 2024.
4. [4] L. Perron, F. Didier. CP-SAT. Google OR-Tools.
5. [5] Tetris Guideline. Super Rotation System (SRS) wall-kick tables.

Cite this case study

@misc{genzero2026tetris,
  title        = {The Tetris Proof: Real-Time Neurosymbolic Superiority},
  author       = {{Gen-Zero Research Team}},
  year         = {2026},
  month        = sep,
  howpublished = {\url{https://gen-zero.ai/cases/tetris.html}},
  note         = {Case Study 01. Reference symbolic controller measured live in the
                  reader's browser; LLM baselines not benchmarked}
}