Define what a test answers
Regression checks verify supported workflows. They do not establish search rankings, performance on every device or a market-wide comparison. Public upstream model tests and ToolkitRack measurements must be labeled separately.
File identification
Use files with known content, retaining, removing and changing their extensions. Compare the predicted type, confidence and extension match indicator. Keep empty, unsupported and low-confidence results in the record. Confidence is not a file safety score.
Compression and conversion
Use owned or clearly licensed samples. Record input format, dimensions, byte size, settings, output format and output size. Reopen results to check dimensions, transparency and visual quality. Include failures and size increases. Per-file byte reduction is (input bytes − output bytes) / input bytes.
SVG tracing
Check for real paths and a viewBox without embedded raster images. Compare edges, colors, transparent areas and small text. Record path count, output bytes and unsuitable examples. An SVG extension alone is not evidence of better image quality.
Reproducible records
Record the date, version, operating system, browser, hardware, settings and cold/warm cache state. Measure page loading separately from processing-to-result time. State the limits of small samples or single-device tests; do not claim “fastest”, “best” or guaranteed outcomes.