Release Notes ============= Mixle 0.8.1 is the credibility, correctness, and evidence release, published as the first release of the 0.8 line (0.8.0 was prepared under the same evidence but never published; decision D-0212). It retains the library's broad probabilistic-modeling surface while making maturity, compatibility, numerical behavior, release claims, and operational limits explicit and machine-checkable. This page describes the 0.8.1 release branch. It is not a publication claim. Final artifact hashes, exact-tip CI results, independent review, and post-publication verification remain release gates. Highlights ---------- Evidence-backed product boundary Public APIs and maturity tiers have machine-readable manifests and drift tests. Stable, provisional, and experimental surfaces now carry different compatibility expectations. Architecture, ownership, security, scientific-validity, migration, and release documents define the supported boundary without presenting research prototypes as production evidence. Numerical and statistical hardening The release includes a broad correctness pass over fused EM, HMM decoding and smoothing, probabilistic-programming transforms, posterior summaries, uncertainty methods, automatic structure learning, categorical support, entropy implementations, and impossible-observation handling. Regression tests preserve each repaired behavior. Reproducible artifacts and execution Safe serialization, schema manifests, provenance receipts, deterministic hashing, checkpoint metadata, base-install import sweeps, minimum-version jobs, platform lanes, and artifact fingerprinting make it possible to tie a claim to the code and artifact that produced it. Distributed training and estimation Existing Spark, Dask, Ray, MPI, and multiprocessing estimation paths have explicit support levels. Packed language-model training gains typed parallel plans, native PyTorch execution where supported, checkpointable training state, and adapters for external distributed-training systems. Multi-GPU performance remains unverified until retained hardware receipts exist; unsupported topology combinations fail before launch. Scientific workflow building blocks The provisional task, reasoning, DOE, causal, calibration, posterior, and lifecycle layers now expose clearer typed contracts and receipts. They can support model selection, distillation, sequential design, evidence-aware decisions, and heterogeneous-data workflows, but do not turn Mixle into a frontier-model trainer or a safety guarantee. Compatibility and dependency changes ------------------------------------ 0.8.1 (as the 0.8.0 candidate) tightens several public contracts in ways that are source-incompatible: they raise, or return something with a different meaning, rather than warning. The classes of change are code-execution gates that now require the ``True`` singleton rather than a truthy value, public Boolean flags that no longer coerce, durable receipt and result records that are now frozen, verification receipts that must name their subject, and a set of numerical corrections that move recorded values. :doc:`migrations/0.8.0` is the authoritative per-surface list and the migration path for each; read it before upgrading. * Python 3.11 is now the minimum supported runtime. Hosted fast lanes cover Python 3.11 and 3.12 on Linux x86_64 and macOS arm64; the full lane uses Python 3.12. * ``mpmath`` is no longer installed by the base package. Install ``mixle[highprec]`` for the mpmath arbitrary-precision fallback, or the separate ``gmpy2`` extra for its supported high-precision paths. * Stable deprecations warn through the shared deprecation helper and remain available for at least two minor releases. Removed or renamed surfaces must have migration documentation and compatibility tests. * Public API additions and removals require a reviewed manifest diff. Names under ``mixle.experimental`` remain explicitly outside the compatibility guarantee. Correctness fixes in the final integration ------------------------------------------ The final stabilization includes fixes for array-backed TreeHMM encoder equality, runtime-resolvable annotations, callable arity probing that could invoke user code twice or mask its exception, deterministic reductions and corpus shingles, explicit optimization failures in place of removable assertions, and early rejection of detailed mixture scores where a scalar objective is required. Validation and remaining gates ------------------------------ Every pull request runs lint, minimum-version, Linux, macOS, clean-wheel, full, documentation, and applicable security checks. The active release branch also runs tests, strict documentation validation, and security auditing on its exact tip after integration. Before publication, the 0.8.1 checklist still requires an immutable candidate artifact, clean-wheel and resolver evidence tied to that candidate, realistic backend and performance receipts for retained claims, ten AI adversarial reviews of the notebook corpus and the examples (replacing the external-tester gate, D-0212), README and documentation stale-content reviews, the final version row, release-owner sign-off, and post-publication verification. See :doc:`release-readiness`, :doc:`claim-evidence-ledger`, and the tracked 0.8.1 checklist for the current state.