assay — limitations

Four geometry families is a ceiling, not a gap

assay ships exactly four families — stroke-grid-monogram, orbital-mark, cut-square, lattice-crucible — and has no plan for a fifth. The spec calls for few knobs and strong defaults: an opinionated generator, not a logo-maker with an ever-growing menu. If a brand needs a shape none of the four can honestly express, assay is the wrong tool, not an unfinished one.

Raster determinism is per-environment, not universal

SVG, tokens.css, tokens.json and BRAND.md are byte-identical across any two runs on any machine, and that is checked directly. Rasterised artifacts — favicon PNGs, favicon.ico, proofsheet.png, og.png — are only asserted identical within one recorded environment fingerprint: platform, OS release, arch, Node version, resvg version.

Font hinting and antialiasing genuinely differ across platforms and resvg releases. assay records the fingerprint that produced a given raster hash rather than pretending a cross-platform byte match that does not exist, and assay diff reports pure and raster drift separately for exactly that reason.

It systematises; it does not invent brand judgment

Palette, font and geometry-family choices still come from a human — or, in the case study, from what a real project already shipped. assay makes those decisions reproducible, contrast-checked and multi-format. It does not decide whether a palette is a good one beyond the WCAG maths it actually runs. Taste in, taste out.

No image model, by design — not by omission

Every visual artifact is computed geometry: SVG path data from pure functions of explicit parameters, rasterised deterministically. There is no image model anywhere in this pipeline and there will not be. That is the reason artifacts can be diffed, version-controlled and regression-tested at all.

Also worth knowing

bundled fonts	Three OSS families (Space Grotesk, Manrope, JetBrains Mono), not a Google Fonts catalogue. Adding one means vetting licence and small-size legibility.
og.png text	Laid out with a character-width estimate rather than real glyph metrics — deliberately conservative, so text never overflows the canvas at the cost of wrapping earlier than a pixel-perfect measurement would.
case-study scope	Four projects from one build program. Evidence that the tool holds up against real, independently designed palettes — not a claim about brand systems at large.
geometry choices	The families picked for each case-study project are assay's own fresh picks, never reproductions of that project's real shipped mark. Each DIFF.md says so explicitly.