What Is New In 0.6.2

Version 0.6.2 broadens Mixle from a composable statistical modeling library into a more complete runtime for heterogeneous modeling, local reasoning, and auditable deployment. The stable center is still the distribution/estimator contract and optimize. The new work adds higher-level creation, reproducibility, placement, reasoning, telemetry, neural, and task-distillation surfaces around that contract.

Use this page as a map of the release. The topic guides linked from each section explain the workflow details.

Local Reasoning Runtime

0.6.2 adds a local application layer for building retrieval, reasoning, and decision workflows around fitted models:

  • mixle.substrate stores text, records, vectors, graph facts, mounted tools, and provenance-bearing evidence.

  • mixle.substrate.Reasoner and investigate plan over retrieve, compute, simulate, and delegate actions while preserving abstention and citations.

  • mixle.inference.skill registers reusable typed capabilities that a reasoner or application can discover.

  • mixle.pool provides a small pool/job abstraction for local or remote execution decisions.

  • mixle.telemetry records decisions, cost, latency, and outcomes; the dashboard helpers summarize those receipts.

  • mixle.scientist assembles the local pieces into an optional, offline scientific-assistant workflow when the scientist extra and local model weights are available.

Read Local Reasoning Ecosystem, Reasoning Systems, and Production Workflows.

Certified Creation, Simulation, And Reproducibility

The inference layer now includes higher-level artifact workflows in addition to ordinary fitting:

  • create(data, ...) fits a model, records provenance, optional calibration, uncertainty, exchangeability diagnostics, and a certificate.

  • simulate(model, ...) draws from models that expose compatible sampler behavior.

  • synthesize(source, ...) creates checked synthetic artifacts from a model, dataset, callable, or task surface.

  • record_fit and verify_reproducible store and check reproducibility receipts.

  • certify and plan_placement expose fit guarantees and local/pool placement decisions.

  • uq dispatches uncertainty queries across fitted models, point predictors, ensembles, and LLM-style callables.

  • hierarchical_event_study estimates confirmed-exposure influence with within-subject shifts, random-effects pooling, difference-in-differences contrast, and sensitivity bounds.

Read Inference, Quickstart, Uncertainty, and Model Lifecycle.

Data, Structure, And Process Families

0.6.2 adds several modeling and diagnostics surfaces:

  • exchangeability_check reports whether rows look exchangeable, shifted, or trended before a workflow assumes that “more rows like these” is a valid sampling story.

  • mixle.represent.modality supplies deterministic image and signal feature helpers for cross-modal examples and tests.

  • Ontology and OntologyConstrainedKG add typed graph constraints for knowledge-graph and reasoning workflows.

  • ContinuousTimeMarkovChainDistribution and its estimator model fully observed CTMC trajectories with a closed-form generator MLE that certifies as GLOBAL_UNIQUE.

  • Multivariate Gaussian fitting uses a BLAS-backed covariance accumulation path and a robust Cholesky fallback so float32/GPU EM fits can recover from small numerical indefiniteness instead of crashing.

  • Hidden Markov distributions default to the Numba encoder when Numba is installed, matching the estimator default while still respecting use_numba=False.

Read Data Layer, Representation Layer, Reasoning Systems, Temporal And Stochastic Processes, Structured Statistical Families, and HMMs and Latent Structure.

Neural And Task Surfaces

The neural and task layers gained more durable, explicit behavior:

  • Flow, MAF, VAE, and DiscreteAR are constructible neural density families that can appear directly in model trees.

  • Neural leaves, direct language models, streaming Transformer leaves, DPO models, energy models, and density leaves now have broader serialization support.

  • Streaming Transformer and DPO accumulators preserve sample weights and EM responsibilities during gradient updates.

  • mixle.task.distill_methods adds Torch-to-Torch response, multi-teacher, hint, attention-transfer, relational, and sequence-level distillation helpers.

  • Task artifact durability improved for JSON-safe qhat=inf reloads, empty inputs, and int4 packed quantized weights.

Read Neural and LLM Models, Task Distillation, and Task Serving, Routing, And Edge Deployment.

Engines, Placement, And Operational Hardening

0.6.2 also tightens lower-level execution and production behavior:

  • Torch DTensor component sharding is gated to Torch 2.5 or newer, where the needed DTensor strategies exist. Older Torch versions get a clear error that points to backend="model_parallel".

  • DTensor import handling recognizes both the public and older private Torch module locations.

  • The production registry constrains names, versions, and aliases to safe path components and raises clearer errors for unknown model names or versions.

  • Benchmark and distributed-stress harnesses that are useful for maintainers moved out of tracked examples/ paths so the examples page focuses on runnable documentation.

Read Compute Engines, Utilities And Parallelism, Examples, and Production Workflows.

API Reference Coverage

The generated API reference now includes the new top-level packages and modules added in this release, including mixle.substrate, mixle.pool, mixle.telemetry, mixle.scientist, mixle.inference.create, mixle.inference.simulate, mixle.inference.synthesize, mixle.inference.reproduce, mixle.inference.skill, mixle.inference.placement, mixle.inference.planning, mixle.inference.orchestration, mixle.inference.event_study, mixle.inference.uq, mixle.data.exchangeability, mixle.represent.modality, mixle.reason.ontology, mixle.stats.processes.ctmc, mixle.models.neural_families, and mixle.task.distill_methods.