v0.7.0a1 semantic limitations¶
This alpha is correctness infrastructure, not deployment qualification or a positive scientific result.
- Exact executable semantics are bounded to the packaged
Q_SCM4:1scalar fragment and frozen corpora; tests and SymPy checks are not formal proofs. - Mandatory tensor support is elementwise NumPy/PyTorch. Broad reductions,
sparse behavior, GPU/JAX parity, and zero-skipping semantics are deferred.
Qualified
ResultTensorarithmetic is currently restricted to the registered CPU float64 execution profiles. Fixed-width integer and complex tensor arithmetic fails before execution rather than wrapping or being mislabeled as an exactQ_SCM4computation. - Bundle schema v3 has a typed ONNX export/runtime path tested locally with CPU ONNX Runtime. This is regression evidence, not qualification. Other execution providers and cross-runtime preservation remain unqualified. Bundle schemas v1/v2 are explicit legacy contracts.
- The first refinement profiles are CPU float64 and operator-bounded. A decoded
result names separate exact denotation, payload origin, and floating
execution evidence.
profile_idis the exact denotational profile (Q_SCM4:1);execution_evidencebinds theFP_SCM4[...]floating profile and its SHA-256. Neither stands in for the other. Asserted scalar results may carry canonicalfractions.Fractionpayloads without inventing floating evidence; the exact oracle remains a separate bounded validation tool. - An exactly-zero denominator is established only by
ExactZeroEvidence.PROVEN_ZEROwith a certificate. The Python type of a zero payload carries no exactness, and exact-zero evidence is deliberately independent of the observed machine payload, soPROVEN_NONZEROwith a0.0payload is a recordable underflow rather than a contradiction. - Exact oracle and production runtime are deliberately separate. Agreement on bounded corpora does not establish universal conformance.
- RF research protocol artifacts are prepared but not authorized. No smoke, pilot, untouched-decision, confirmatory, robotics, compiler, or deployment experiment result is claimed by the alpha.
- External algebra, methodology, and formal novelty review remain pending.
- Working
ResultTensorobjects are not durable audit records. Freeze them toResultTensorSnapshotbefore serialization, monitoring, or external handoff. - Projective array-like conversion is bounded to 10,000,000 input elements and 32 nested sequence levels before NumPy/Torch conversion. Larger carriers are rejected rather than partially converted.