Falken-9
A resident neural tic-tac-toe model designed for the Commodore 64. It powers Crystal Palace 9—the existing game and artwork, now with sound and the Falken-9 model. The 5,195-byte model deployment packet retains attention, learned routing and two selected experts per move. Every AI mark executes the network; there is no minimax or board-to-move lookup hidden in the game.
Emulator-tested only. Physical C64, 1541 and SID compatibility and performance are unverified. This is a custom integer runtime artifact, not a Transformers checkpoint, GGUF, Ollama model or hosted inference endpoint.
Actual graphical011 emulator output, not a mockup. Crystal Palace 9 is the game;
Falken-9 is its model. Earlier falcon-9 / FALCON-9 labels were a
historical spelling inconsistency, not a different model. The model name and repository slug are
Falken-9 and falken-9.
Downloads
| File | Purpose |
|---|---|
| model.f9j | Frozen F9J2 Q11/W16/A32 deployment packet; 5,195 bytes |
| C64 PRG | Standalone game with resident model, graphics, original SID music and effects |
| C64 D64 | Disk image containing that same standalone game |
- Game guide and screenshots — controls, loading, opening variety, audio, limitations and acceptance boundaries.
- Illustrated research report — exhaustive findings through study016, 39 selected indexed figures including the original weight and connection views, worked arithmetic, controls and counterexamples.
- GitHub source and probes — assembly, training/verification tools, test scripts, research probes and authoritative art.
- Project website — browser launch helpers, disk inspection, downloads, game screenshots, music preview and HTML report.
- Model description · Release manifest · SHA-256 checksums
No code or probe archive is required merely to run the supplied PRG/D64. The separate model packet is for the custom Falken-9 runtime described in the source repository. Downloading it does not make it compatible with a general LLM loader.
The 5,195 bytes refer only to the model deployment packet; the complete game, 6502 inference runtime, graphics/audio and working memory require additional space.
What the model computes
Nine board-square tokens encode occupancy, absolute position and side to move. The width-8 integer network performs normalization, full-board attention, projection/residual, flattened readout, learned routing, two of four expert MLPs (hidden width 4), a merge and the final nine-score head.
A nonnegative score indicates membership in the complete optimal-move set; score magnitude is not calibrated confidence or a categorical preference. Fixed code validates history, canonicalizes the board, runs inference, restores the whole approved set to original coordinates, masks occupied squares and applies the declared tie priority. Offline minimax supplies training/audit labels, never gameplay moves.
The frozen packet is model-v1.0.0, SHA-256:
fe180f8a931f9d92d8aa39da337bc819d0dbe4e6d971edcba00324fbc18a91c1.
All 627 canonical boards pass the established assembled numerical checks. The 4,520-raw-board wrapper result is a composed verification, not 4,520 separately executed native games. Recent interpretability, opponent and robustness studies are host analyses with separately stated evidence boundaries.
Why this model exists
Across Palace-9, Crystal-9 and Falken-9, the engineering question is:
Can a learned attention-and-expert network execute on a C64, play correctly, and be small enough to investigate exhaustively?
Beating a specialized tic-tac-toe solver on speed is not the acceptance criterion. The experiment concerns real learned computation on a constrained machine, verified gameplay and a model small enough for systematic investigation.
The progression was Palace-9 (standard local-LLM/GGUF tooling), Crystal-9 (a leaner custom model), and Crystal Palace / CP64 (the original Crystal-9 model executing on a C64 with disk paging). Falken-9 instead trains for C64 arithmetic, live RAM and work per turn from the start. These are distinct models, not interchangeable formats or successive exports of the same weights. Users seeking GGUF/llama.cpp/Ollama should see Palace-9.
The complete model stays resident, with no per-turn model-data disk loads. It powers Crystal Palace 9, retaining the game's artwork while adding sound and a faster neural execution path. In informal emulator observations, complete AI-vs-AI games lasted just over a minute, with roughly ten-second thinking intervals and visible activity: a quick curiosity to watch rather than a prolonged wait. An earlier model-backed version took roughly 20 minutes. These are user-observed emulator impressions across releases, not a controlled benchmark or a physical-hardware speed claim.
Decoded weights and tensor workflow
The original weight-map image, preserved byte-for-byte. Each cell represents one decoded scalar parameter, not a bit or byte. Colors normalize per tensor; all four expert banks are displayed, while only two selected experts execute per inference. Static weights are not board-dependent activations.
What the research establishes
The network has context-dependent causal effects, but no isolated “fork expert” or complete symbolic strategy circuit has been established. The report separates learned computation from wrapper-provided symmetry, stored weights from dynamic activity, full approved sets from selected ties, and individual from joint parameter robustness. It also documents self-play's limited coverage, the fixed-seed sampler's nonuniform opening frequencies, and exact outcomes against a specifically defined random legal opponent. Self-play is not learning.
Inspiration, credits and license
Inspired by WarGames (1983), Falken-9 is named for Dr. Stephen Falken, the fictional researcher who developed Joshua, the AI on the WOPR war-simulation computer. This project plays tic-tac-toe only; its war-room setting is fictional. “Patient Calculations” is an original SID score, not a film melody.
Project by Lewis Moten, with Mercury Quill / Hermes Agent assistance. First-party material is Apache-2.0 to the extent the contributors hold the relevant rights; applicable artwork provenance and limitations are in NOTICE and LICENSE. No Commodore ROM, emulator, assembler, third-party soundtrack or installed software environment is redistributed here.