Changelog#

The repository’s CHANGELOG.md is authoritative — this page mirrors it.

All notable changes to mononet will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased#

Added#

  • mononet.legacy — a deprecated, backend-agnostic (Keras 3) drop-in port of the original airtai/monotonic-nn MonoDense layer, its create_type_1/create_type_2 builders, and helper functions. Emits a DeprecationWarning on use; keeps the original {-1, 0, 1} monotonicity indicator. Serves as a migration bridge for existing airtai/monotonic-nn users. Numeric equivalence to the original is pinned by committed goldens.

  • Public package skeleton with mononet.core, mononet.torch, mononet.jax, mononet.keras layers implementing the constrained monotonic construction.

  • MonotonicityMask, ActivationSpec, InitSpec, MonoConfig framework-agnostic value objects in mononet.core.

  • NumPy reference function signatures pinned by tests.

  • Cross-backend equivalence test directory (tests/equivalence/) ready for the future harness.

  • Four devcontainer flavors: default (CPU) + gpu-torch, gpu-jax, gpu-keras (CUDA 12.4 base, Python 3.13).

  • CI matrix: 3 Python versions × 3 backends on Ubuntu + Python 3.13 on macOS and Windows.

  • PyPI trusted publishing (OIDC) workflow.

  • Sphinx (myst-nb) site rewrite with guides, concepts, benchmarks, and about sections; benchmark notebooks committed with their outputs.

  • NOTICE.md with patent reservation + commercial-license contact.

  • tools/execute-benchmarks.sh for manual notebook re-runs before releases.

Changed#

  • BREAKING: the default mode is now "mixed" (was "split") for MonoLinear / MonoDense / MonoResidual / MonoConfig / MonoResidualConfig. mixed uses the static init (no init needed) and is the paper’s base |W| construction — pass mode="split" explicitly to keep the previous behaviour.

  • BREAKING: renamed the mode values "absolute" → "mixed" and "switch" → "split" (MonoLinear / MonoDense / MonoResidual / MonoConfig / MonoResidualConfig); the old names now raise ValueError naming the replacement.

  • BREAKING: MonoLinear / MonoDense now default activation to None, meaning "identity" — a linear monotone map, matching torch.nn.Linear and keras.layers.Dense(activation=None) (MonoConfig keeps a concrete ActivationSpec("identity") default). Layers that relied on the implicit ReLU are now linear — pass activation="relu" explicitly to restore the previous behavior.

  • BREAKING: MonoResidual and MonoResidualConfig now require an explicit activation when the default F is built (a custom F must not also pass activation), preventing a silently-linear residual branch.

  • Layer activation parameters are now typed ActivationSpec | ActivationName (ActivationName = Literal["relu","elu","selu","softplus","identity"]) instead of accepting an arbitrary str, so unknown names are rejected at type-check time.

  • Relicensed from PolyForm Noncommercial License 1.0.0 to the Apache License 2.0, following AIRT Technologies Ltd.’s decision to discontinue patent-related activities. Apache-2.0’s section 3 grants the patent rights needed to use the code. Effective from the first PyPI release.

  • Switched LICENSE from proprietary (cookiecutter default) to PolyForm Noncommercial License 1.0.0 (assignee: AIRT Technologies Ltd.).

  • Python support range broadened from 3.13-only to 3.11–3.13.

  • Removed runtime pydantic dependency; configs use stdlib dataclasses.

Removed#

  • 1Password integration in devcontainer initialization.

  • Legacy private PyPI index registry (synthpop-pkgs) and matching UV_INDEX_SYNTHPOP_PKGS_* workflow secrets.

  • Linear workflow files (.linear.toml, LINEAR_GUIDE.md, the linear-cli Claude skill).

  • Codecov configuration and CI upload step.

  • Second cookiecutter devcontainer flavor (partner).

  • HelloWorld placeholder in mononet/__init__.py.

0.1.0 - 2026-04-13#

Added#

  • Initial release of mononet

  • Basic package structure with Python 3.11+ support

  • Development toolchain with uv, pytest, and pre-commit hooks

  • Comprehensive linting and static analysis setup

  • Documentation framework with Sphinx

  • CI/CD pipeline with GitHub Actions

  • Automated version management and publishing