Scientific data contract

This is the user-facing contract boundary for current development behavior. The semantic model describes the interpretation layer; generated nouns and verbs define specific accepted objects and arguments. Different operations need different dimensions: static terrain, vector roads, station series, summaries, and spatiotemporal rasters do not share one universal cube shape.

Aspect	Implemented checks or propagation	Assumptions and limits
Dimensions	Named reductions require their dimension; canonical raster names are time, y, x, optional band	Adapters may retain native names. Not every callable normalizes them
Coordinates	Overlap requires exact labels; harmless dimension ordering can normalize	Equality does not establish matching observation footprints
CRS	Nouns retain source CRS; overlap preflight rejects known conflicts; spatial operations have specific inference/reprojection rules	Absent CRS is a review condition in semantic reports, not a universal execution error
Temporal meaning	State tracks labels and known support; explicit alignment can require exact support	Missing support stays unknown; daily labels alone do not authorize PRISM/gridMET interchange
Units	State retains known units; variance squares them; conditions/proportions are dimensionless where defined	There is no complete dimensional algebra or automatic cross-source conversion
Metadata	Maintained verbs update result meaning, dimensions, units and definition records	Metadata inference cannot prove source descriptions correct
Provenance	Nouns retain provider/product, native variable, query and source/revision records	Pipe traces do not capture all preparation, branches or downstream work
Missingness	Threshold comparisons and condition means follow documented two-valued rules	No general valid-observation denominator, uncertainty propagation, or minimum-valid threshold

Units and compatibility

PRISM temperature is Celsius; gridMET temperature is kelvin. Converting units is a visible analytical operation, not an automatic benefit of a shared noun. Threshold numbers must use the input unit. Overlap's coordinate/CRS/support checks do not certify compatible thresholds or physically equivalent products. Unknown units remain unknown, and validate() reports metadata conditions without computing values to discover their true physical interpretation.

See condition/reduction semantics, temporal support, and source contracts. The internal spatial contract includes legacy CRS fallbacks; follow each operation's actual API instead of assuming every spatial helper fails on missing CRS or every non-spatial object requires time.

Serialization

Source and continuous-summary scalar metadata are NetCDF-safe under the documented export policy. For condition/state Datasets, use to_netcdf: Boolean variables become flagged int8 in a write-only copy, preserving the in-memory object. Structured metadata is serialized canonically; unsupported custom objects fail rather than yielding unstable representations. A downstream reader must interpret flags explicitly; writing a file does not certify its later scientific use.

The checks are in first-use acceptance tests. This is object serialization, not automatic export of complete workflow lineage.

Where protection stops

Enforced: cataloged input kinds/required information, operation-specific coordinate/support checks, explicit export constraints.

Advisory/inferred: inferred semantic state, unknown metadata checks, order messages, suggestions and explanations. These expose software interpretations without choosing a scientifically correct method.

Planned/unresolved: three-valued conditions, durable merged branch lineage, complete uncertainty propagation, automatic scientific harmonization, and universal semantic rule coverage. Reserved names are not implementation.

A valid Python expression and a successful computation can still answer the wrong question. The analyst must justify source, support, thresholds, weights, missing-value handling and interpretation. Review the evidence matrix for exactly which software claims have test coverage.