Concepts

AI-ready data retain the scientific context needed to interpret and reuse values: units, methods, biological entities, observational support, quality, source rights, and a reproducible transformation history. A vector alone does not supply that context.

Multimodal describes different ways of observing a system, such as field measurements, traits, sensors, fluxes, imagery, gridded products, or text. Multiscale describes spatial and temporal support, such as organism, plot, site, landscape, hourly, seasonal, or annual. Many datasets do not necessarily constitute many modalities. Matching coordinates does not establish equal support.

Core harmonization contains conservative reusable operations: explicit unit conversions, WGS84 coordinate validation, missing-value normalization, or verified identifier mapping. Track harmonization contains application-specific selection, aggregation, neighborhoods, metrics, and interpretation. The contract rejects filtering, joins, and support-changing operations in core recipes. Promotion of a track assumption requires a separately reviewed scientific decision and implementation; there is no automatic promotion path.

Human-governed autonomy means that agents may abstain and propose decisions, QA may reject them, and consequential uncertainty must reach a human. A human approval cannot turn an unknown relationship into an explicit equivalence; revise the mapping with evidence first.

Representations are machine products derived from validated observations. Model preprocessing is recorded separately from scientific harmonization. Modality-specific vectors stay separate; this prototype makes no universal embedding claim.