VISION PRECISION BENCH · PUBLIC PREVIEW · NOT A FULL RANKING YET

Did the model see it — or only sound fluent?

VPB measures fine-grained visual accuracy across six tracks: change localization, text reading, confusable characters, counting, color, and relative size — then composes a transparent VPI.
535frozen cases
6tracks
v0.2.1benchmark version
1transparent VPI

Why multi-track

Not caption fluency

Models can describe a scene while missing micro changes, tiny text, or O vs 0.

Deterministic scores

Geometry, exact strings, integers. LLM-as-judge is not the primary bar.

Versioned freeze

v0.2.0 is a public preview freeze. Claims without a version are not VPB results.

Honest composite

VPI never hides a failed track. Per-track columns are mandatory.

This site is Keel IR

Pages are pure fn → View. Packaging is engine ssr_site_document_html + app_shell HostDocument — the same class of document product React/Solid ship, without a dual product framework.