0.9.0

Highlights

• jax networks export to ONNX for HSSM. Networks trained with jaxtrain now export through jax2onnx (transform-jax-onnx), producing files HSSM loads directly (#109). jax2onnx 0.15 to 0.17 is supported (#130, #150).
• Networks are published where HSSM looks for them. upload-hf keeps the full artifact folder and also writes the ONNX file at the repository root under HSSM's filename, together with a root manifest.json (#110). A malformed architecture or training section in model_card.yaml is now rejected (#134).
• Training runs describe themselves in MLflow. Each run logs the model, network type, backend, parameter bounds, config hash, LANfactory version and a run UUID that links the MLflow run to its files on disk (#108). The learning rate is logged every epoch, and a cosine schedule is available (#124).
• Faster GPU training. The trainer no longer forces a GPU synchronisation on every step (#120).
• Network inspectors. The new lanfactory.network_inspectors package plots what a trained network has learned, including interactive manifold plots (#82, #99).

Fixes

• The trainers work with pandas 3 (#140).
• The jax trainer uses a supplied lr_dict instead of ignoring it (#146).
• A "linear" hidden activation in a jax network no longer shifts every later activation by one layer, which crashed jaxtrain or applied the wrong activation (#153).

Changes to check before upgrading

• An unknown learning-rate schedule name now raises an error instead of being ignored (#124).
• ssm-simulators>=0.15.0 is required (#154). New runtime dependencies: jax2onnx, plotly and seaborn.
• The documentation moved to the shared ecosystem platform and gained a page on what LANs are (#114 to #127).

What's Changed

• Add network_inspectors.py module in LANfactory by @mariama-design in #82
• Add interactive rendering for network inspector manifold plots by @mariama-design in #99
• Make training runs self-describing in MLflow by @AlexanderFengler in #108
• jax → ONNX export via jax2onnx: jaxtrain now produces HSSM-consumable artifacts by @AlexanderFengler in #109
• Scheduled drift detection: weekly full-suite re-run against fresh resolves by @AlexanderFengler in #107
• Publish networks where HSSM actually looks for them (dual HF layout + manifest) by @AlexanderFengler in #110
• docs: phase-0 safety net — fix config rot, strict build in CI, deploy workflow, redirects, link check by @AlexanderFengler in #114
• Correct metadata and docs our own merges made stale by @AlexanderFengler in #112
• docs: adopt the shared ecosystem brand bundle (phase 1) by @AlexanderFengler in #115
• docs: real landing page + network-variant consolidation (phase 2) by @AlexanderFengler in #117
• docs: canonical contract-page pointers + rank-1 overclaim fixes by @AlexanderFengler in #119
• Stop synchronising the GPU on every training step by @AlexanderFengler in #120
• docs: Diátaxis nav restructure + What LANs are explanation page by @AlexanderFengler in #122
• docs: point at the ecosystem hub instead of a partial map by @AlexanderFengler in #123
• Log the learning rate, add cosine, refuse unknown schedules by @AlexanderFengler in #124
• docs: adopt the shared documentation platform by @AlexanderFengler in #125
• Align the documentation legal notice by @AlexanderFengler in #126
• Complete the LANfactory documentation quality closeout by @AlexanderFengler in #127
• chore: ignore generated .cursor/ and .agents/ directories by @AlexanderFengler in #128
• fix(hf): reject a non-mapping architecture/training in model_card.yaml by @AlexanderFengler in #134
• chore(deps-dev): bump mkdocstrings-python from 2.0.5 to 2.0.7 by @dependabot[bot] in #132
• chore(deps-dev): bump mkdocs-redirects from 1.2.2 to 1.2.3 by @dependabot[bot] in #131
• chore(deps): update jax2onnx requirement from ==0.15.* to >=0.15,<0.17 by @dependabot[bot] in #130
• fix(trainers): update training history indexing to use iloc by @cpaniaguam in #140
• Use supplied lr_dict in ModelTrainerJaxMLP instead of ignoring it by @EItanm1999 in #146
• Generate 250 parameter sets in the e2e tests so files divide the batch sizes by @AlexanderFengler in #151
• ci(docs): sync reusable-docs.yml (actions/deploy-pages 5.0.1) by @AlexanderFengler in #152
• chore(deps-dev): bump mkdocstrings-python from 2.0.7 to 2.0.9 by @dependabot[bot] in #148
• chore(deps): update jax2onnx requirement from <0.17,>=0.15 to >=0.15,<0.18 by @dependabot[bot] in #150
• Move linting/formatting check into separate CI workflow by @cpaniaguam in #138
• Refactor test setup with temporary directories by @cpaniaguam in #95
• Migrate to ruff 0.17 and lift the <0.16 cap by @AlexanderFengler in #155
• Raise the ssm-simulators floor to 0.15.0 by @AlexanderFengler in #154
• Release 0.9.0 by @AlexanderFengler in #156
• Keep JAX MLP activations aligned with layers for hidden "linear" by @fmuia in #153

New Contributors

• @mariama-design made their first contribution in #82
• @EItanm1999 made their first contribution in #146
• @fmuia made their first contribution in #153

Full Changelog: 0.8.0...0.9.0