Release notes.

Accuracy amendment. Corrections to clauses whose stated method did not match what the canonical runner measures, plus removal of assertions the calibration data does not support. No parameter was added or removed and no breakpoint was retuned, so scores computed under the previous text remain valid.

MEASUREMENT

• Measurement is performed at the image's own resolution. The earlier text expected input that had already been down-sampled, which made results depend on whoever prepared the file.
• COL-02 scope corrected. Computing the whole-polygon mean across every neutral surface produced roughly five times the scale of the production measurand, so a conforming runner would have scored against mis-scaled thresholds. The scope now defers to the canonical runner.
• Saturation is measured on HSV S over 0 to 255 throughout, with the single carve-out for the set-consistency chroma statistic, which remains defined on CIELAB.

SCORING

• New clause 6.9 introduces evidence-only parameters: measured and reported with status EVIDENCE, carrying no numeric score and excluded from aggregation, with the remaining weights renormalised. COL-03 was previously neither scored nor N/A and had no valid recording state.
• Worked example corrected: the composite arithmetic and the EXP-01 parameter score now reproduce exactly under the clause's own breakpoints.
• Set consistency now notes that with population sigma the maximum attainable z-score is the square root of one less than the set size, so outlier attribution at |z| of 2 or more is arithmetically impossible for sets smaller than six images.

EDITORIAL

• Values deferred to the canonical runner are now deferred consistently, rather than being stated in one clause and deferred in another.
• The corpus is described as the June 2026 production corpus rather than an accepted-output corpus, since it contains images carrying production flags.

First release of the VQS Specification: the measurement framework, the parameter set across six categories, set consistency, the scoring model, the result record schema, and the certification and governance clauses.

PUBLISHED

• Measurement framework: colour management, working representations, region classes, normalised coordinates and reproducibility requirements.
• Parameter set across Exposure, White Balance, Colour, Geometry, Artifacts and Semantic plausibility, each with its measurand, direction of merit and applicability rules.
• Scoring model: continuous breakpoint mapping, the discrete severity rubric of 100, 80, 55 and 20, N/A exclusion with weight renormalisation, category scores and the composite.
• Bands of PASS at 90 and above, FLAG from 70 to 89 and FAIL at 69 and below, aligned to the certification tiers.
• Set consistency statistics as properties of a listing set, explicitly not blended into any per-image composite.

CALIBRATED

• Thresholds calibrated against real production QA output rather than estimates, including the warm-asymmetric white balance anchor band, the geometry bounds and the saturation ceilings.
• Content that could not be supported by that data was removed rather than published as provisional, so the document carries only verified values.