QA Sub-Agent with Visual Verification

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

A slide contains a long two-line title and an image caption. Structural validation can pass while substituted fonts clip the title or move the caption over the image. Actual rendering needs inspection.

Mechanism

Record final file identity, renderer, fonts and delivery dimensions, then render every page. Inspect title bounds, overlap, footers, contrast and reading order. An independent QA context can reduce author expectations when authorized. After source fixes, rerender changed and shared-template-affected pages and match delivery bytes to evidence. Web artifacts also need interaction checks.

Bad example

Inspect slide XML coordinates and claim no clipping because all numbers are inside the page, without opening renders.

Good example

Render every slide at delivery dimensions and inspect long title, caption and footer. Record visible clipping/overlap, fix and rerender affected pages. Retain renderer and file identity and state unverified interactions, fonts or viewers.

Why the change matters

Structure does not capture all font, wrapping and display behavior. Renders expose visible results; final-version binding prevents earlier images from becoming delivery evidence after edits.

Observable expectation

The teaching checklist has a render and observation for every page. Narrowing a title box should expose clipping; a new render after repair confirms completeness. A changed final hash invalidates old images. Unopened screenshots cannot support no observed defects.

Limits

Visual QA supplements structure and accessibility, not animation, links or all viewers. Alternate renderers may handle fonts differently; record limits. A passing screenshot does not establish native target-application behavior.