Known boundaries
Evidence aid, not an oracle.
The workbench makes visual review more structured. It cannot determine whether a changed region is correct, why a defect occurred, or whether an interface is ready to ship.
It does not judge language
No screenshot comparator can reliably assess meaning, grammar, register, terminology, plural selection, gender, cultural sensitivity, or legal wording. Missing or duplicated-looking strings still need source context and a qualified reviewer.
It does not certify accessibility
A screenshot does not expose semantic HTML, accessible names, relationships, focus order, keyboard traps, screen-reader announcements, zoom behaviour, reflow beyond the captured viewport, or input alternatives. Colour observations also need measurement in the running product.
Pixel differences are not defect decisions
- Font rasterization and fallback can change many pixels without a localization defect.
- Animation, caret blink, loading state, dynamic data, and experiments create unstable captures.
- Image compression and colour-profile conversion add noise.
- A threshold can hide subtle clipping or emphasize harmless antialiasing.
- Different dimensions prevent a reliable one-to-one overlay.
Only captured states are covered
A clean image says nothing about another viewport, browser, theme, validation state, user role, content length, input method, or locale. Full-page captures may miss sticky, virtualized, lazy-loaded, or interaction-dependent content.
Local-first trade-offs
There is no built-in account, shared workspace, team permission system, remote backup, issue-tracker synchronization, or cross-device history. Browser storage is not durable archival storage and may be cleared. Exports require the reviewer to manage access, retention, and versioning.
Project files are checked for corruption, not for tampering
An exported .lqaw project carries an FNV-1a-32 integrity value over its canonical payload, and the workbench refuses a file whose value does not match. That check detects accidental corruption in transfer or storage. It is not a cryptographic signature and not tamper evidence: anyone who edits a project file can recompute the value, so a matching project file does not prove who produced it or that nobody changed it. Treat project files the way you treat any other unsigned working file.
Performance and format boundaries
Very large images can exceed device memory or browser canvas limits. The workbench pauses automatic pixel diff above a device-sensitive memory threshold while keeping side-by-side, swipe, overlay, blink, region, and pixel-loupe review available. A reviewer can request one larger diff after an explicit warning, but the browser may still stop that analysis. Decode support differs by browser. The documented workflow focuses on still PNG, JPEG, and WebP screenshots; other image, document, design, or video formats are outside the input contract.
No automatic root-cause analysis
The workbench does not inspect source code, translation catalogs, CSS, DOM direction, font loading, or build configuration. A finding should be reproduced in the running interface and investigated by the responsible engineering, design, localization, or accessibility team.
Does zero highlighted difference mean the locale is correct?
No. It may mean the images are identical, the wrong files were paired, the threshold hides the change, or a hard-coded source string remained untranslated.
Can the report be used as compliance evidence?
It may support a documented review, but the workbench makes no compliance or certification claim. Use the standards, experts, and formal procedures required by your organization.
Will local data always survive?
No. Browsers can evict or clear site data, and devices can fail. Export important work and store it in an approved repository.