Format

The .ndtbl format stores one multidimensional field group in a compact binary layout with self-describing metadata followed by one contiguous payload block.

On-disk encoding

All integers and floating-point values are encoded in little-endian byte order. Floating-point values use IEEE-754 binary32 and binary64.

Header

Each file begins with this fixed header:

• magic[8]: "NDTBL\0\0\0"

• version:u8: currently 1

• scalar_type:u8: 1=float32, 2=float64

• reserved:u16: must be zero

• payload_offset:u64: byte offset of the first payload value

• dimension:u64: number of axes

• field_count:u64: number of named fields per grid point

• point_count:u64: total number of grid points

Axis records

One axis record is stored for each axis in axis order:

• kind:u8: 1=uniform, 2=explicit_coordinates

• reserved0:u8: must be zero

• reserved1:u16: must be zero

• extent:u64: number of points on the axis

Uniform axes append:

• min:f64

• max:f64

Explicit-coordinate axes append:

• extent repeated f64 coordinates

Field names

Field names are written in payload order as:

• name_length:u64

• name_bytes[name_length] encoded as UTF-8

Payload ordering

The payload stores raw scalar values in point-major order.

Logical point traversal follows row-major / C-order over the grid axes:

• axis 0 is the slowest-varying grid index

• the last axis is the fastest-varying grid index

• for each logical point, all field values are written consecutively in field_names order

Equivalently, the payload is the flattened form of an array shaped as (*axis_sizes, field_count) in C-order.

Validation rules

Readers should reject files when:

• reserved fields are nonzero

• point_count does not match the product of axis extents

• payload_offset does not match the parsed metadata length

• encoded sizes exceed supported host limits