mixle 0.6.2

Broadens mixle from a composable statistical-modeling library into a runtime for heterogeneous modeling, local reasoning, and auditable deployment. The stable center remains the distribution/estimator contract and optimize.

Highlights:

• Local reasoning runtime (mixle.substrate): retrieval, multi-hop reasoning, grounding/factuality, and governance around fitted models.
• Telemetry surfaces and production placement/reproducibility helpers in mixle.inference.
• Constructible neural-density families: VAE, Flow, MAF, DiscreteAR (no build_* + adapter double-wrap).
• Broader knowledge distillation (response/dark-knowledge, multi-teacher, hint/FitNets, attention-transfer, relational, sequence-level).
• Correctness fixes: neural / DPO / StreamingTransformer accumulators now honor per-item weights; registry paths constrained to the store root; structure-search estimator cloning uses deepcopy instead of eval.

See the "What Is New In 0.6.2" guide in the docs for the full map.

mixle 0.6.1

Highlights since 0.6.0.

Neural densities as composable distributions. Normalizing flows (coupling and masked-autoregressive), a VAE, mixture density networks, autoregressive-categorical densities over discrete vectors, and an energy-based model — each a mixle Distribution that fits jointly with classical families by EM and drops into the mixle.ppl surface (Flow, MAF, VAE, MDN, EBM, CondFlow, ...). The mixle.models neural leaves were also renamed off the ...Leaf suffix (back-compat aliases kept).

Automatic dependency-structure learning. optimize(data) / fit(data) now discovers the cross-field dependency graph by default (BIC-gated, never worse than independence). Heterogeneous Bayesian networks with regression edges, mixtures of DAGs, and counterfactual queries on the learned graph.

Cross-modal reasoning (mixle.reason). Product-of-experts and hybrid attention/structured fusion, a transfer-safe adapter over frozen encoders, iterated-EKF assimilation of nonlinear forward models, and retrieval by what the model believes.

Task distillation (mixle.task). solve() for regression, multi-label, and structured routines with conformal calibration; a generative-text student; agent-trace harvesting; grammar-constrained plan generation; and live-traffic health monitoring.

Inference & tooling. Closed-form Fisher-weighted parameter merge, closed-form variational projections (exact GMM collapse + Runnalls reduction), exact multi-prime NTT convolution for count-DP ranking, and a refreshed documentation site.

---

• Install: pip install mixle
• Docs: https://gmboquet.github.io/mixle/
• Notebooks: https://github.com/gmboquet/mixle-notebooks

mixle 0.6.0

First release under the mixle name (formerly pysparkplug / pysp). pip install mixle; import mixle.

Automatic inference for composable models of heterogeneous data — a model and the estimator that fits it have the same shape, and fit chooses the inference (conjugate / EM / MAP / variational / MCMC) from the model itself.

Highlights:

• PPL language core: deterministic-expression slots (Normal(a + b, sd)), potential() custom log-factors, .each() / .each(by=...) random effects, non-Normal GLMMs (Poisson/Bernoulli(... + Group)), data-indexed latents (theta[Field("g")]), and per-parameter R̂/ESS in summary().
• Lighter fitting API: optimize/fit accept a distribution prototype or just the data (estimator inferred); fit forwards every optimize knob.
• Categorical(free) / Dirichlet(free) infer their dimension from the data; unified streaming estimators.

Originally developed at Lawrence Livermore National Laboratory (LLNL-CODE-844837).