# Vectorize independent logo-vectorization evaluation protocol

Status: protocol frozen on September 16, 2026. No independent results have been published yet.

## Purpose

This protocol defines the next evaluation before its source logos are selected or processed. It exists to prevent the current five-family development benchmark from being presented as independent evidence.

## Separation from the development set

- Adidas, Slack, Wikimedia, Starbucks and Google are excluded.
- No evaluation source may be used to tune segmentation, palette selection, tracing or cleanup before the run is complete.
- The runner commit, settings and rendering environment are recorded before the evaluation assets are processed.
- A failed case remains in the report. It is not replaced after results are known.

## Source eligibility

Each source must have a public provenance page, a stable download URL and a recorded SHA-256 hash. The set must cover distinct problems rather than near-duplicates:

- one-color lettering and counters;
- separated multicolor regions;
- circles or symmetric geometry;
- thin internal detail;
- a wordmark at low resolution;
- transparency and a white-on-transparent case;
- a JPEG input with an opaque background.

Reference vectors are used only to create controlled raster inputs and to measure rendered output. They are never passed to the tracing pipeline.

## Frozen input variants

Where licensing and source geometry permit, each family is evaluated as:

1. a clean large raster;
2. a clean low-resolution raster;
3. one realistic degradation selected before processing, such as JPEG compression, an opaque background or mild shading.

The exact dimensions, compression settings and background colors are recorded with the source manifest.

## Measures

The report keeps measurements separate:

- rendered silhouette overlap;
- mean and minimum color-region overlap;
- connected components and holes;
- palette count and color difference;
- traced region count and output bytes;
- processing time;
- a visual review of lettering, symmetry, small gaps and detached artifacts.

No composite “quality score” is produced. High overlap does not override a topology failure or visibly incorrect lettering.

## Publication rule

The complete set, settings, metrics, output artifacts and failures are published together. The report identifies the run as independent only if the sources were unseen during development and no tuning occurred after inspection of their results.

This document intentionally describes the evaluation method without distributing Vectorize implementation code.