Numerical source types

Measurement arrays retain their original type until validation. Numeric text, booleans, datetime values and duration values are refused before conversion to float64, including booleans and durations inside object arrays. Integer and floating arrays remain supported. Where an API already accepts numeric object arrays, every original element must be a plain real number. Shape, finiteness, index, probability and domain bounds remain owned by the individual API. Coupling prior, spectral coupling, transfer entropy and delayed dynamics refuse real numeric object values outside the float64 range with the owning input error before native dispatch.

This contract applies to coupling, phase extraction, monitor inputs, replay policy arrays, imprint updates, SSGF, supervisor distributions and gradients, phase gossip, delayed dynamics, plasma coupling imports, PHA-C acceptance evidence and numerical visualisations. For hybrid quantum monitoring, classical phases obey this rule; the quantum state retains its complex statevector or density-matrix contract.

Recurrence outputs are binary masks and accept boolean arrays as well as valid numeric 0/1 arrays. Other original measurement types remain refused. Generated activation masks, random fixtures, histograms and validated numerical results are internal calculations with their own types.

Language boundaries

Go, Julia and Mojo Python bridges share the same original-source checks before packing numerical buffers. Their kernels receive typed numerical buffers. Rust measurement and metadata extractors reject source aliases before conversion. The named-call wrappers retain positional, keyword and mixed calls and their existing introspection signatures. Negative unsigned counts retain the existing OverflowError. Plain measurement/metadata extractors raise ValueError for source aliases; incompatible ndarray ABI types are rejected by PyO3 type checks.

The flat hypergraph encoding contains one start offset and strength per edge, without a terminal offset. Offsets start at zero and strictly increase; each edge contains at least two distinct in-range nodes. Invalid offsets, node indices, buffer dimensions and non-finite values raise Python errors before integration. Empty native pairwise coupling and alpha buffers preserve their existing meanings: absent pairwise coupling and zero phase shifts. A zero-step run validates its inputs and returns the initial state.

Measurements

Delayed dynamics and transfer entropy comparisons record the subsequent overflow and diagnostic corrections against all five available backends, with the corrected source hashes.

Final language comparisons record real calls to the available Python, Rust, Go, Julia and Mojo backends. Parity gates compare numerical results; the separate trajectory timing rows measure elapsed time and do not establish numerical parity. Records include fixture arguments and the installed extension hash. Earlier dated measurements remain historical observations of their recorded source hashes.

The host is shared. These small fixtures establish input and numerical contracts; they do not establish production scaling or controlled speedup ratios.