Release Notes¶
The 0.7.0 release is a capability and hardening release. It expands Mixle’s task, DOE, neural, latent-model, and reasoning surfaces while tightening numerical behavior in mixture, automatic-inference, and registry paths.
Highlights¶
This release focuses on three themes:
stronger numerical behavior for mixture, EM, HMM, automatic-inference, and missing-data paths;
broader task, distillation, DOE, neural, and reasoning APIs; and
clearer documentation for maturity, validation, production use, and family release expectations.
Added¶
- Task and distillation capabilities
The task layer now includes richer teacher/student workflows: soft-label distillation, structured task distillation, active labeling, cascade economics, local harvest/retrain loops, and agentic task distillation for tool calls and plans.
- DOE for distillation and cross-modal training
mixle.doe.distillationadds pool-based experiment design for task distillation and cross-modal training. It helps choose informative teacher calls, balance task coverage, and treat label acquisition as an expensive experimental design problem. See Design of Experiments for the Sphinx examples and selector contract.- Neural and energy models
The branch adds reusable neural-model builders including Deep Sets, monotonic MLPs, input-convex networks, Hamiltonian networks, and energy-based product-of-experts helpers.
- Latent and dependence models
New or expanded latent surfaces include gated mixtures, copulas, structured mixture/reduction utilities, and additional mixture-of-experts style building blocks.
- Reasoning and evidence surfaces
The reasoning package adds cross-modal transport checks, task-sufficient projections, cycle-consistency signals, anchor harnesses, answer receipts, and provenance-aware explanation helpers.
Changed¶
- Automatic modeling is more defensive
Empty data, all-empty nested sequences, detector failure paths, marginal field validation, and model-recommendation fallbacks now fail more explicitly.
- Mixture and EM paths are harder to destabilize
Mixture code has additional stress coverage for high-dimensional Gaussian mixtures, singular or near-singular covariance paths, weighted responsibilities, and impossible-observation updates. See Numerical Stability and Missing Data for the branch-level contract around
NaNinputs,-infimpossible observations, robust mixture initialization, and DOE score validation.- PPL route behavior is more explicit
The PPL guide now documents
explain_fit, explicit missing-data marginalization, composite custom potentials, state-space fitted distributions, and indexed latent sampler routes. Unsupported route/feature combinations are expected to raise clear errors instead of returning partially applied models.- Task documentation is broader
The task guides now cover one-call replacement patterns, calibrated structured outputs, density gates, serving cascades, economics, extraction, and agentic/task-planning variants.
Fixed¶
The branch includes hardening commits for foundational EM, precision, registry, automatic-inference, oracle timeout, stochastic test, safetensors, and mixture-stability issues. See the repository history for the exact patch commits used for the final tag.
Compatibility And Migration¶
No broad removal is documented for this release. Users upgrading from the previous release should pay closest attention to expanded task APIs, DOE distillation helpers, reasoning receipts, and the clarified missing-data and impossible-observation contracts.
Validation Focus¶
The release should be validated through the gates in Release Readiness. At minimum, release evidence should cover:
python -m buildandtwine check dist/*;install from the built wheel in a fresh virtual environment;
import sweep over public
mixlemodules with optional-dependency guards;full test suite, not only the fast marker subset;
examples and notebooks that are shipped or linked by the docs;
strict Sphinx build with warnings as errors; and
the coordinated family resolver/integration check across sibling packages.