METHODOLOGY

Reliable by construction

Versioning

SemVer for the public claim surface. v0.2.0 freezes 511 cases. Changing metrics or cases bumps the version — silent patch freezes are forbidden.

Scoring

Primary metrics are deterministic (IoU match, exact string, integers). VPI is a disclosed weighted mean; missing tracks make VPI null.

Contamination

Synthetic seed-locked SSOT for public preview. Real annotated packs are planned for 1.0 under the same schemas.

Scale policy

~500 multi-track cases matches Lite/Verified-style public boards: enough cells for curves, cheap enough for multi-model runs. 0.3 targets 1k–1.5k; 1.0 targets ≥2k + real images.