Numeric data representation

Every analytical method in the product reads and writes one shape of numeric data: a packed double[], described by a row count and a column count, rather than a data frame, a DataTable or a CSV.

• Packing version 1 is IEEE 754 binary64, little endian, row major. For a shape of Rows by Columns, the packed byte length must equal Rows * Columns * 8; a caller cannot hand over a six value array for a 3x4 (twelve value) shape.
• The packed array is the one interchange format between a data source, a training run and a trained model’s recall: nothing downstream re-reads a data frame, a file or a live database connection once a dataset has been packed.
• An expired, foreign or malformed pointer to a packed array reads as not found rather than throwing a parsing error.

This is the structural counterpart to a data frame in R: it has a narrower job (one numeric matrix, no column names, no mixed types) and no row-recycling behaviour — a shape mismatch is always rejected, never silently repeated to fit.