Problem Statement

Media-processing systems often report that they completed a task without independently proving that the produced image or video is acceptable.

For privacy processing, a pipeline may blur most frames but miss a few. For masking or redaction, the intended area may remain visible. A global image metric can hide these local failures, while manual frame review is slow and difficult to reproduce.

Visual Verifier addresses a narrower problem:

Given original media and a processed candidate, measure meaningful visual changes and return reproducible evidence about whether the current verification rule passed.

Scope of the engine

Anonymization is the named use case because that is where a missed frame is most expensive, and it is what people search for. The engine itself is general: it measures pixel-level change between a reference and a candidate and makes no assumption about why the pixels changed. Watermarking, overlay application, transcoding, filter chains, compositing, and ordinary processed-media regression tests all use the same contract.

Current scope

The active package compares aligned media with matching content. It detects changed regions, filters weak regions, measures severity, and evaluates a generic presence-of-processing rule.

Desired properties

• Deterministic local execution
• Typed Python and CLI interfaces
• Machine-readable results
• Human-reviewable evidence
• Explicit failure codes
• Reproducible thresholds
• No dependence on archived notebook state

Non-goals of the current release

• Performing the transformation
• Identifying semantic objects
• Proving privacy compliance
• Correcting alignment automatically
• Replacing human review in high-risk workflows