Tic-tac-toe raw-history and symmetry corpus

A deterministic, generated corpus for 3×3 tic-tac-toe move-policy research. It contains the chronology-preserving, nonterminal legal histories used by the custom raw-history contracts documented in Crystal-9 and Palace-9, plus reproducible board-state and symmetry-reduced views.

This is a dataset repository. It contains no model weights, runtime, checkpoint, or inference service.

Square alphabet and board representation

Each letter identifies a board square. A history is the occupied-square letters in chronological order: odd positions are X moves and even positions are O moves.

Letter	Row	Column	Grid
a	1	1	top-left
b	1	2	top-center
c	1	3	top-right
d	2	1	middle-left
e	2	2	center
f	2	3	middle-right
g	3	1	bottom-left
h	3	2	bottom-center
i	3	3	bottom-right

 a | b | c
---+---+---
 d | e | f
---+---+---
 g | h | i

For example, history ae means X occupies a, then O occupies e:

X . .
. O .
. . .

JSONL board uses slash-separated rows (X../.O./...) and board_text uses the written three-line grid above. . means empty.

Labels and terminal semantics

optimal_move is one deterministic minimax target. It is the first optimal legal move using fixed tie order:

e b d f h c g i a

optimal_moves retains every equally optimal move in that same order. This lets consumers distinguish the deterministic target used by the released policies from other equally correct continuations.

There are no win labels and no winning-line annotations in this data. The rows describe moves, boards, and policy targets only. A sequence that tries to continue after a terminal board is invalid; terminality is represented only by the lack of a legal continuation, never by a supplied winner or line-of-three field.

! is reserved by the related model contracts for malformed, repeated-square, terminal, or otherwise out-of-protocol input. It is not a next-move target in nonterminal-legal-histories.jsonl.

Contents

File	Rows	Purpose
data/nonterminal-legal-histories.jsonl	294,778	Chronological legal histories through eight moves with deterministic and all-optimal targets. Primary raw-history corpus.
data/reachable-board-states.jsonl	5,478	Unique boards reachable by legal alternating play, including positions where play cannot continue.
data/nonterminal-board-states.jsonl	4,520	Unique reachable boards that have a legal continuation.
data/rotation-canonical-reachable-board-states.jsonl	1,383	Unique reachable boards after canonicalization across the four rotations.
data/dihedral-canonical-reachable-board-states.jsonl	765	Unique reachable boards after canonicalization across rotations and horizontal reflection (the 8 square symmetries).
data/manifest.json	—	Generation contract, exact row counts, and SHA-256 for every JSONL artifact.
generate_dataset.py	—	Standard-library-only deterministic generator.

State-space statistics

Every square has three raw possibilities (., X, O), giving $3^9 = 19,683$ board patterns. A pattern is invalid here when no legal alternating move chronology reaches it without continuing after terminal play.

View	All ternary boards	Reachable boards	Invalid board patterns
Raw board space	19,683	5,478	14,205
Rotation-canonical board space	4,995	1,383	—
Rotation-and-reflection canonical board space	2,862	765	—

The rotation and dihedral outputs are deterministic transform views that reduce duplicate geometric board states. The primary history corpus deliberately keeps chronological move order because the Crystal-9 and Palace-9 interfaces accept histories, not canonical board encodings. Use a canonical board file only when your own training/input contract explicitly applies the matching transform to both board features and move coordinates.

Reproduce exactly

No network access or third-party package is required:

python3 generate_dataset.py --output data
sha256sum -c SHA256SUMS

The generator implements the move alphabet, legal-history walk, terminal-continuation rejection, minimax evaluation, fixed tie order, board-text rendering, rotations, and reflections. Regenerating data/ must reproduce its manifest.json row counts and artifact digests.

Provenance and scope

This is synthetic game data generated by exhaustive deterministic rules. It contains no collected gameplay, personal data, external game records, or claims about general gameplay intelligence.

• Crystal-9 is a custom local runtime move-policy model. Its published exhaustive policy gate covers 294,778 nonterminal legal histories.
• Palace-9 is a raw-history Qwen2-MoE model with GGUF compatibility.

The datasets make shared game semantics inspectable; they do not replace either repository's release-specific training, validation, or runtime evidence.

Project lineage

This dataset continues Lewis Moten’s long-running exploration of tic-tac-toe implementations and game AI. Earlier public projects include:

• Tic-Tac-Toe (2014) — offline-first HTML5/JavaScript game and Newell–Simon-inspired AI exploration.
• Tic-Tac-Toe LSL (2015) — interactive Second Life game with graduated AI difficulty.
• Tic-Tac-Toe AI (2016)
• Artificial Neural Network (2016) — neural-network learning project with game experiments.
• Octothorpe X/O (2016) — reusable JavaScript tic-tac-toe game engine.

These projects are historical context, not training-data sources or dependencies of this generated corpus, Crystal-9, or Palace-9.

Acknowledgments

This dataset was designed and directed by Lewis Moten. Its code and documentation were developed with assistance from GPT-5.6-terra Med, accessed through Hermes and using Honcho for context and project-memory support. Lewis Moten remains the project designer, maintainer, and publisher.

License

Apache License 2.0. See LICENSE.