Units that differ by a thousand, buffers that live in another file, quantized vertices, and a node hierarchy 24 levels deep. What to check before a GLB hits the GPU.
Arabic has six plural forms, Polish has four, and most i18n bugs are a missing one. Eleven formats, the plural rules that break them, and what RTL actually requires.
Obfuscated fonts, an NCX that disagrees with the nav document, vertical Japanese, and a spine pointing at a file that is not there. The ebook cases worth owning.
Tokenisation that drops the term someone typed, ranking that buries the exact match, and zero results with no way out. The failures a search box hides.
Nested MIME, four charsets in one mailbox, authentication headers that pass and fail, and a bounce that looks like a reply. Test your parser against real message shapes.
Variable axes, COLR palettes, metrics that disagree, and a deliberately broken checksum. The font cases that break renderers, subsetters and export pipelines.
Antimeridian crossings, unnormalised longitudes, silent CRS assumptions, and a shapefile that is really five files. Test the cases that break importers.
Extractors fail on codecs they never linked, filenames they cannot decode, and nesting they never expected. Here is how to find out before your users do.
Assemble a compact editing regression pack from six-role audio and video workflow kits, expected analyses, controlled failures, and local browser-generated controls.
Turn a creator brief into a repeatable AI workflow test with provider-neutral plans, bound media inputs, expected outputs, and an explicit evaluation rubric.
Build synthetic media fixtures from versioned recipes, understand byte-stable versus content-stable exports, and verify every artifact without uploading source media.
Use one canonical field contract to test browser forms, paged layouts, structured values, field-data exchange, escaping, and round-trip imports across six form families.