快乐掼蛋 — 高级 AI 模型与第三方组件 DanZero V1T INT8 weights: Author/project: dashidhy / DanLM https://github.com/dashidhy/DanLM Pinned revision: d200e455f63fd46aa398e97b69a6bb94cf1ba098 Original artifact: ckpts/DanZero_v3_rep_v1t/v3_rep_v1t_best_eval_001_int8.onnx Weight file is unchanged. This site independently implements a JavaScript adapter for its own rules, public observations, legal actions, endgame checks, and explanatory UI. This is not the larger DanLM Transformer. Full terms: DanLM-LICENSE.txt (Apache-style terms with Section 10, Non-Commercial Use Only). This free family game uses it non-commercially. Commercial redistribution/use requires the permission specified there. ONNX Runtime Web 1.31.1, Microsoft Corporation: https://github.com/microsoft/onnxruntime Unmodified WASM runtime and JavaScript distribution files. License: onnxruntime-LICENSE.txt Third party notices: onnxruntime-ThirdPartyNotices.txt Inference happens locally in a Web Worker. The model receives only the acting player's cards and public table events. No hidden opponent cards are supplied. Generated explanations describe observable tradeoffs; model scores are not calibrated win probabilities or private thoughts. Local self-play calibration tunes two bounded decision preferences. It does not update neural-network weights. Old game saves and legacy training records remain in their original browser storage.