The following is the upstream model notice. The source in source/Mortal corresponds to the cited commit. The platform-specific native binary is obtained separately at first startup.

Mortal runtime source
Upstream: https://github.com/Equim-chan/Mortal
Commit: 0cff2b52982be5b1163aa9a62fb01f03ce91e0d2
License: AGPL-3.0-or-later (LICENSE-MORTAL.txt)

This distribution includes the upstream libriichi and exe-wrapper source,
Cargo manifests and lockfile, and mortal/model.py and mortal/engine.py.
Upstream files retain their original content. The release-specific inference
and build scripts are distributed under AGPL-3.0-or-later.

Model weights: Apache-2.0 (https://www.apache.org/licenses/LICENSE-2.0).

2026-11-03: This local match kit adds a manzu/pinzu two-view ensemble,
enables the upstream agari guard for hanchan, and includes a reproducible
comparison harness and regression tests. "S42+" names this inference profile,
not an upstream model release. The weights and upstream source are unchanged.
The added inference, server, launcher, terminal, benchmark and test code is
distributed under AGPL-3.0-or-later; see LICENSE-MORTAL.txt.

GitHub distribution: the original ONNX model is downloaded and SHA-256
verified on first launch, pinned to revision
2f7ad48bc29dadd9805b1897f3c5b97b76c0d41a of
https://huggingface.co/xzhe121/Mortal-S42-onnx.
The relevant Mortal runtime source is included in source/Mortal. Upstream
documentation and log-viewer artwork are omitted from this GitHub copy;
the complete upstream source remains available at the commit cited above.
The comparison summaries are included in evaluation; the original full
game logs are part of the earlier ZIP distribution and are reproducible
using benchmark.py with the recorded seeds.

2026-11-04: Muse用アルゴリズム v1 is a local derivative. Self-distillation
changes 47,149 advantage-head weight/bias scalars; the encoder and value row
remain fixed. The learned head is in weights/muse-head.npz and the original
head is in weights/base-head.npz. prepare-model.py applies the learned head
to the pinned base and verifies the complete derived-model SHA-256.
Both the UI and the default inference profile use the derived model. The
48-hanchan comparison does not establish a playing-strength gain. Training
code is AGPL-3.0-or-later; derivative weight tensors retain Apache-2.0.
See training/report.json and 追加学習と検証.txt for provenance and limits.
Earlier dated statements refer to the previous version only.



2026-11-04 (Q4 distribution package): This package omits the local
source/Mortal copy to reduce size. The relevant Mortal runtime source
remains available from the upstream repository at the commit cited above
(0cff2b52982be5b1163aa9a62fb01f03ce91e0d2). The AI model is distributed
as the 4-bit compact pack (compact-models/model-q4.mqp); the FP32
model.onnx is not included. Decision agreement of the Q4 pack with the
FP32 model was 98.8% over 2,304 non-forced positions from FP32 self-play.
