Frozen results, stated with scope.

Visual metrology · V7

The V7 real-world screenshot cohort and release regression report provide the following validated results:

• Integer translation component error: 0/0/0 px median/p90/max for 16 returned estimates.
• Subpixel Euclidean error: 0.018 px median and 0.039 px p90 for five returned fractional estimates.
• Zero false-confident vectors on 22 no-change controls. Separately, all six controlled ambiguous-texture cases abstained.
• Top-3 changed-area localization: 22/22; zero false-positive regions in top three.
• Comparison p95: 0.618 s for the V7 benchmark cohort.

Source: Artifacts/V7/REPORT.md, frozen V7 cohort. The subpixel sample is five estimates; one real translation abstained. These cohort figures do not predict every image.

Codex client compatibility · Interface Hardening V1

The frozen suite used 32 unchanged, blinded tasks in fresh Codex sessions against the updated live MCP executable:

• Tool selection: 32/32 (100%).
• Valid MCP arguments: 32/32 (100%).
• Abstentions respected: 7/7 (100%); no fabricated measurements after unavailable results.
• Unit/status interpretation: 32/32 (100%).
• Accurate final explanations: 31/32 (96.9%).

Source: Artifacts/InterfaceHardeningV1/REPORT.md and its frozen compatibility results. This validates a specific Codex CLI study, not all clients or future versions.

Current source artifacts

The website copy maps to the frozen repository artifacts: Artifacts/V7/REPORT.md, Artifacts/GeneralVisualMetrologyHardeningV1/REPORT.md, and Artifacts/InterfaceHardeningV1/REPORT.md. No example demo measurements are presented as benchmark data.