How we test, and when a page’s date changes

Source priority

When a claim can collide with marketing language, we prefer evidence in this order:

• Primary documentation: official platform help, API docs, standards (C2PA, EXIF, ISO-BMFF), and model cards.
• Direct observation: behavior we can reproduce from a file in a current browser, or from the tool’s own output.
• Attributed third-party measurements: a named benchmark with version preserved.
• Community reports: useful, but labeled anecdotal until we can reproduce them.

Browser and format support

Metadata Cleaner accepts JPG, JPEG, PNG, WEBP, MP4 and MOV with a parseable ISO-BMFF/QuickTime box structure. HEIC/HEIF is not implemented. Safari, Chrome, Firefox and Edge are in scope for the cleaner’s Canvas / FileReader path; very large 4K masters can still exhaust a tab.

Local OCR uses Tesseract.js in the browser with English, Simplified and Traditional Chinese, Japanese, Korean, Spanish, French, German, Portuguese, Hindi and Arabic. Packs lazy-load after Run OCR. The Auto option does not detect script; it falls back to English. A UI language is not an OCR pack. Accuracy is published in the 2026 multilingual OCR benchmark on controlled fixtures, not as a single marketing percentage.

Video Compressor requires WebCodecs. There is no WASM fallback. Hardware acceleration is preferred when the browser exposes it; we do not claim a specific encoder was used unless the browser says so. A size preset is a target, not a guarantee.

GIF to Live Photo encoding and pairing/structure validators are browser-tested. Apple Photos LIVE recognition, Samsung Gallery Motion Photo recognition, and Android/Google Photos import are not yet verified on a physical device. Lock Screen wallpaper eligibility is not guaranteed.

PDF Toolkit copies page objects for merge/split/extract/reorder/rotate/remove where pdf-lib allows. PDF → Images is a rasterizing path and is labeled as such. Encrypted PDFs are not cracked.

Provenance evidence grades

The AI Image Origin Detector separates local inspection from optional official OpenAI verification.

• Trusted positive — an official OpenAI C2PA/SynthID result after the user chose to verify, or a structurally present C2PA/JUMBF packet the local inspector can show. Still not a statement about pixels.
• Useful clue — EXIF, XMP, PNG text, software tags, filenames. Easy to strip or rewrite.
• Missing / inconclusive — this scan did not find the field. Absence is not a human-origin proof. Screenshots and social recompression often destroy file-level credentials.

OCR accuracy

OCR quality depends on resolution, compression, language choice, and layout. Receipt totals, handwritten notes, and tiny UI type fail often. We do not publish a single “accuracy %” for marketing.

Searchable PDF export overlays Tesseract word boxes. Force OCR on a native-text PDF can leave both the original text layer and the OCR layer.

Empirical benchmarks and zero-fabrication research (/labs/)

Our empirical laboratory (/labs/) publishes repeatable, open benchmarks derived directly from repository test fixtures and client runs. We enforce a strict zero-fabrication policy: we never estimate, smooth, or market fake accuracy percentages.

For example, the 2026 Multilingual OCR Benchmark evaluates browser-native Tesseract.js 5.1.1 across 10 languages and real document types (thermal receipts, invoices, identity cards, packaging, menus). Every published CER, WER, and key-field accuracy metric is calculated from docs/benchmarks/ocr/2026-09/results.json and validated in CI via tests/ocr-benchmark-engine.test.mjs.

Complete raw datasets and fixture manifests are publicly downloadable in JSON and CSV formats under /labs/ocr-benchmark/dataset/ so anyone can audit and reproduce our findings locally.

Video compression notes

Platform file-size limits (Discord, LINE, WhatsApp, and others) are copied from official docs when we cite a number, with a verification date on the article. We do not treat a competitor’s blog as the source of a limit.

Compression settings change size and quality; they do not control what a platform does after upload.

Privacy wording

A “no upload” sentence must match the default path of that tool. Optional OpenAI verification, remote URL fetch, optional model download, and AI Text Detector (remote TypeSafe System One, with OrcaVerify Text 1.0 as fallback) are disclosed on those screens and on About.

AI Text Detector is not local-only. Submitted writing leaves the device.

We do not log private filenames or file contents for analytics.

When dateModified changes

We update dateModified, article:modified_time and sitemap lastmod when the URL’s unique content materially changes: a platform limit, a product behavior, new evidence, or a substantial rewrite.

We do not bump those dates for CSS, header/footer chrome, analytics, or a one-letter spelling fix.

Sitemap lastmod is taken from this registry and from each article’s own dates. It is not the build timestamp and not the git checkout mtime.