The Stable Surface¶

This is the reviewed, deliberately short list of what 0.8.2 supports as stable (worklist A1.3): the surfaces covered by the compatibility policy in Support Policy, whose behavior is pinned by tests and whose changes follow the deprecation lifecycle. It is short on purpose – short enough to test exhaustively and read on one page. Everything not on this list is provisional or experimental per the machine registry in mixle.maturity; a whole namespace is never declared stable when only a subset is mature.

The rule: a surface is stable only with artifact-level evidence – a test or gate that fails if the behavior regresses. Each row below names that evidence.

Stable surface	Evidence (fails on regression)
Gaussian, Gamma, Exponential, LogGaussian, Poisson, Geometric, and Categorical distribution modules	the complete invariant catalog for this allowlist, their per-family interface and recovery suites, and `scipy_golden_test` density parity
`mixle.stats.parameter_packing`	exact pack/unpack shape, ordering, and round-trip contracts
`mixle.semantics` (see Quantitative semantics)	the packaged `fixtures/quantitative-semantics-v1.json` cross-project contract round-tripping without loss, the semantic-identity boundary suite (operational fields – sample location, backend, job – excluded from identity while priors, units, transforms, constraints and observations are not), and the refusal of any `schema_version` other than the supported one
Direct MLE / EM / conjugate fitting through `optimize`	the weighted-estimation contract (weights == replicated sufficient statistics), the fit-seed determinism suite, and EM monotonicity/quiet-by-default behavior
Base NumPy execution (no optional backend installed)	the blocking clean-wheel install + import-sweep job and the base-install optional-import guard
Serialization paths explicitly covered by compatibility tests	the five explicitly stable persistence schemas (Gaussian, Poisson, Exponential, Categorical, and nested Mixture), their digest-bound 0.7.0 load fixtures and `0->1` migrations, the dependency-profile schema drift gate, and atomic-write / safe-JSON deployment tests. Other registered types are versioned but provisionally persistent.

Not stable¶

Everything else is explicitly not covered by the stable compatibility promise, even where it is useful and tested:

provisional (usable, may change within a minor release) – all other mixle.stats families (including combinators, mixtures, HMMs, graphs, processes, and Bayesian catalogs), mixle.ppl, mixle.process, mixle.models (neural leaves, GPs, grammars, …), mixle.task / mixle.reason, mixle.enumeration / mixle.ops beyond the capability-gated core, mixle.doe / mixle.evolve, the runtime layers (mixle.substrate / mixle.pool / mixle.telemetry / mixle.scientist), and mixle.inference.production (see Project Maturity);
experimental (no compatibility guarantee) – everything under mixle.experimental and the standalone frontier-training mechanism prototypes (muP, 2:4 sparsity, scaling laws, simulated TP/PP/CP, and fault injection). The executable distributed backend and complete-state checkpoint APIs are provisional mixle.utils.parallel surfaces, not stable compatibility promises.

To check a surface’s tier programmatically:

from mixle.maturity import maturity_of
maturity_of("mixle.stats.univariate.continuous.gaussian")  # -> Maturity.STABLE
maturity_of("mixle.stats.latent.hidden_markov")            # -> Maturity.PROVISIONAL
maturity_of("mixle.ppl")                                   # -> Maturity.PROVISIONAL

The stable list and the machine registry are kept consistent by a test, so this page cannot silently claim more than the registry backs.