Claim-Evidence Ledger
=====================

Every material public claim about mixle, with the strongest evidence that currently supports it. The
point of a credibility release is that claims are graded, not asserted: a reader can see exactly how
far each statement has been verified and decide whether that is enough for their use.

Evidence grades (from the 0.8.0 release contract):

- **E0 — assertion:** prose, design intent, or a unit test with a stand-in only.
- **E1 — local correctness:** a deterministic unit/integration test on one machine.
- **E2 — artifact correctness:** a test from a clean installation of the built wheel.
- **E3 — realistic workload:** real public data, or a real external model/backend, at bounded scale.
- **E4 — independent reproduction:** another environment or person reproduces it from documented steps.
- **E5 — production/scale:** sustained real deployment, or representative multi-node/multi-GPU runs.

Stable-core claims target at least **E2**; performance and backend claims target at least **E3**.
"Production-ready", "safe", or "frontier" claims require **E5** and are **not** made in 0.8.2. Grades
here are the release owner's current assessment and are revised as evidence lands (several are pending
the exact-tip re-run for this release, marked below).

.. list-table::
   :header-rows: 1
   :widths: 30 12 58

   * - Claim
     - Grade
     - Evidence / caveat
   * - ~90 distribution families with a common estimator/sampler/encoder contract
     - E1
     - Source-tree family, sampler-seed, and scipy-golden density tests cover the broad provisional
       catalog. A clean-wheel import sweep is not behavioral evidence for every family.
   * - Seven stable scalar families satisfy normalization, score, fit, and serialization invariants
     - E2
     - The exact wheel is installed into an isolated environment; the stable invariant catalog runs
       there with no source-tree import path and retains JUnit evidence.
   * - "15,000+ tests"
     - E1
     - Measured 2026-09-08: 16,849 collected (``pytest --collect-only -m ""``) on the 0.8.2 candidate.
       The tiers run in CI (``core`` on 3.11–3.12 and both platforms, ``full`` sharded on 3.12,
       ``optional`` sharded); not yet run from the wheel for the whole suite (only the import sweep
       is) — that remains an exit criterion for this release too, unchanged from 0.8.1.
   * - One ``optimize(...)`` call fits a heterogeneous composed base model
     - E2
     - The collection-light public fit path runs from the isolated installed wheel and retains JUnit
       evidence. Neural and latent compositions remain E1 unless their optional installed-wheel lane
       is retained separately.
   * - A PyTorch module fits in one line with parity to a hand-written loop
     - E1
     - ``torch_parity_test`` / ``grad_control_test`` check parity, freezing, and checkpointing (torch extra).
   * - Distillation into a small local model with calibrated deferral and a cost/quality receipt
     - E1
     - Task/calibration/cascade tests are E1 on synthetic teachers. The repository carries no direct
       dataset usage (2026-08-04 decision), so real-data task evidence lives in notebooks outside it
       and is not claimed here.
   * - HMM / GMM performance and crossover behavior
     - none
     - The harness is tracked in ``benchmarks/``. Older measurements are archived and are not release
       evidence; numbers may be published only after an exact-candidate retained run.
   * - Distributed estimation over Spark / Dask / Ray / MPI by switching one argument
     - E1
     - Per-backend encoded-data tests exist. Dask and torchrun are exercised in the scheduled optional lane;
       Spark, Ray, MPI, and Lightning remain tested locally but not CI-gated. Wording is qualified in the
       backend support matrix.
   * - Distributed transformer and MoE training
     - E1
     - The native backend executes DDP/FSDP2/HSDP and selected TP/CP paths; Megatron Bridge owns full
       TP/PP/CP/EP/ETP. CPU objective/checkpoint and provider-delegation tests establish local correctness,
       not a realistic-scale run. GPU throughput, MFU, and multi-node recovery remain unverified until
       retained hardware receipts exist. Unsupported axes fail capability validation.
   * - Serialization round-trips and provenance / replay
     - E1
     - Fresh-process serialization tests; lineage/receipt tamper-and-replay tests. Cross-version
       (load a prior release's artifact) fixtures: digest-bound 0.7.0 load fixtures for the stable
       persistence schemas exist (see :doc:`stable-surface`); broader cross-version coverage remains a
       0.8.1 gate (worklist M11.2).
   * - Calibrated abstention (conformal / cascade / escalation)
     - E1
     - Conformal-coverage and calibration-diagnostic tests. This is a deferral mechanism, not a safety
       guarantee — no "safe to deploy" claim is made (needs E5).

This ledger is maintained alongside the release checklist; when a claim's evidence changes, update the
grade in the same change.