Vectors and tiles

Scarp v0.0.5-alpha.1 includes numerical features built on finite binary64 values. These are specialized primitives, not general-purpose mutable collections or tensors.

Vectors

A vector is an immutable collection of finite floats. The compiler example shows vector literals, indexing, snapshots, and numerical kernels:

bind values = [1.0, 2.0, 3.0, 4.0];
print(values, len(values), values[0]);
var weights = [0.2, 0.3, 0.5];
bind snapshot = weights;
weights = scale(weights, 2.0);
print(snapshot);
print(weights);
print(sum(values), mean(values));

len, scale, sum, and mean are implemented numerical built-ins. Other checked operations include dot products, norms, reductions, windows and selected prediction/loss utilities. See built-ins rather than assuming Python/NumPy function names exist.

Tiles and matrix operations

A tile has two explicit dimensions with immutable row-major numerical elements.

bind weights = tile([1.0, 2.0, -1.0, 3.0], 2, 2);
print(matvec(weights, [2.0, 1.0])); // [4.0, 1.0]

The currently implemented tile API includes height, width, cell, lane, unroll, turn, matvec, matmul, tileadd, and tilescale. There is no automatic broadcasting, mutable cell indexing, or general N-dimensional tensor type.

Build a real two-axis numerical example

The verified matrix-forward lesson reproduces the complete published matrix_forward.srp file and its expected stdout. Unlike a short illustrative snippet, its source/output pair is included in the documentation's pinned compiler conformance suite and executed in CI. Start with verified vectors if vector is unfamiliar.

matvec(t, v) requires width(t) == len(v); matmul(a, b) requires width(a) == height(b). A tile's cell count is at most 65,536, with row/column dimensions checked separately; row-major elements cannot be changed in place. Each new materialized output can charge the cumulative --max-vector-bytes budget, so successful type checks do not guarantee unlimited computation. Consult numerical diagnostics before increasing limits.

Classification and probability kernels

The verified classification and probabilities tutorial runs three full published compiler examples covering matrix scores, argmax tie-breaking, softmax, crossentropy and onehot targets. All source and expected output blocks are checked against the pinned release, including a separate compiler execution job. These examples demonstrate numerical primitives, not a training system.

Newly published alpha.2 statistics and training primitives

The alpha.2 collections and numeric-learning tutorial provides four additional compiler-verified complete programs: immutable integer collections, stable score selection, numerically stable probabilities, and explicit loss gradients. The separate alpha.2 fixture suite checks them against the newer release; the existing alpha.1 numerical examples remain independently pinned. These primitives do not provide automatic differentiation or a full training loop.

Scientific computing roadmap (unreleased)

Scarp is extending numerical work incrementally. The scientific computing proposal distinguishes shipped vector/tile operations from proposed integration, differentiation, tensor and AI work. The sampled-calculus functions in compiler PR #153 are not in the published alpha.3 compiler; do not treat them as runnable released APIs.

Limits and errors

Indexing is bounds-checked. Invalid shapes, nonfinite results, and numerical payload-budget exhaustion produce diagnostics rather than silently returning invalid data. Each vector/tile element is a finite 64-bit float; some operations materialize new buffers and consume the cumulative vector-byte budget.

Verified against: vector reference, tile reference, numerical fixture.