Project: Noesis

Contextual compute, certainty scoring, pattern-oracle interfaces, and fast tactical support layers.

What Noesis Is

Noesis is broader than the GOAM game supplements. At the systems level, it defines contextual compute operators, certainty scoring interfaces, and pattern-oracle machinery that can operate in deterministic or entropy-enabled regimes.

GOAM is one application line: there, Noesis appears as the fast tactical and certainty-aware layer used for quick reads, cached patterns, and instinctive support. The game artifacts on this page are therefore application evidence, not the full definition of Noesis.

The clearest proof is tic-tac-toe. In the GOAM supplementary artifacts, the Noesis tic-tac-toe cache is only about 289 KB and reaches forced optimal play with roughly 700 learned patterns. That compactness matters because it shows the approach can become decisive long before brute-force search becomes necessary.

Chess is the larger stress test, not a solved-domain claim. The supplementary chess artifact uses the same pattern-cache logic but at a much larger scale, making the right framing "instinctive quick play and certainty support" rather than exhaustive completion.

Self-Play Behavior

Against itself, Noesis should trend toward an even match. Small self-play slices can still show a temporary X/O split because tic-tac-toe gives first move initiative, and the runs were sampled under nonzero temperature rather than hard deterministic play. That is not the same thing as saying the pattern oracle is biased.

An early tic-tac-toe training run at temperature = 1.2 logged 50 games, 375 total moves, 93 cache hits, and a 26/21/3 split between X, O, and draws. The plateau detector still marked the system as "still learning at game 50," which is the right way to read that snapshot: it is a pre-convergence training view, not the mature forced-play result.

The training bias toward center and corner openings helped self-play converge faster, which is exactly what you want in a compact domain like tic-tac-toe: learn the decisive structures early, then force optimal continuation from a small cache.

Chess Snapshot

There is an important caveat on the chess side: the earliest encoder/self-play passes were pre-validation experiments. Move legality was not being enforced correctly yet, which means some of those runs were effectively debugging-era "anarchychess" and should not be presented as evidence for Noesis performance.

Chess behaves differently. A representative 10-game self-play snapshot logged 1772 total moves, 419 cache hits, 47737 cache misses, and a result spread of 2 white wins, 0 black wins, and 8 draws at the sampled stopping point. That should not be over-read as a stable white advantage. In longer stopping windows, the pattern tends to alternate toward small white/black win counts with the majority of games resolving as draws.

That draw-heavy profile is the real signal. Noesis is not brute-forcing solved chess. It is reaching high-certainty, mostly stable continuations often enough that self-play spends most of its time preserving balance, with occasional tactical breaks when one side's sampled continuation drifts harder than the other's.

The temperature split also matters here: one logged run used white_temp = 0.6 and black_temp = 1.2, so the point of the snapshot is to show pattern behavior under noisy play, not to claim a final equilibrium result.

Pattern Keying

For chess, the pattern cache had to become more explicit than the tic-tac-toe board string. The cache entries are keyed so they stay both machine-usable and legible to a human reviewer: effectively a hashed identifier tied to FEN + side-to-move + move, with the stored record preserving the board state and tactical flags.

{
  "0056992a41fe7b3940d5cb4202ef682d_w_b1c3": {
    "fen": "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1",
    "move": "b1c3",
    "certainty": 3.3616087436676025,
    "outcome": "draw",
    "move_number": 1,
    "side_to_move": "white",
    "castling_rights": "KQkq",
    "en_passant": "-",
    "is_check": false,
    "is_capture": false,
    "is_castling": false
  }
}

API Reproducibility

The Noesis scoring service is exposed through https://entropy.occybyte.com. The important architectural point is that scoring is geometric at the API layer rather than heavyweight ML inference. That keeps reproduction lightweight: no local GPU stack is required to verify certainty scores or cached pattern behavior.

The API supports single-move and batch certainty scoring, direct pattern evaluation, and cache queries. Batch scoring is the preferred path because it reduces call overhead when evaluating multiple candidate moves from the same state.

The same service shape is already generalized beyond chess and tic-tac-toe. The documented state encodings include chess, tic-tac-toe, Farkle, Leduc Poker, and blackjack, which makes Noesis look less like a one-off game heuristic and more like a reusable pattern-oracle interface.

Supplementary Artifacts

These artifacts currently live as GOAM supplement materials rather than standalone public downloads on this site.

Associated Technical Paper

GOAM is the main paper. Noesis appears there as a supporting tactical and certainty-aware layer rather than the sole subject of the publication.