Vision calibration: how these numbers were earned

Primary data behind Mart402's vision products. Everything below comes from measurements we ran ourselves; the dataset design and error classes are published so you can judge the products - and so the field has a citable reference for what vision-language models actually get wrong about food.

meal-vision: 30 weighed meals

We calibrated against 30 dietitian-portioned meals that were weighed on a scale (Japanese teishoku, donburi, noodle bowls, Chinese, Western). Protocol: blind - the model's estimate is fixed and recorded before the true weights are revealed.

• Median kcal error: about 12-15 percent across the set
• The disclosed two-sided kcal range contained the true value in 26 of 30 meals
• Every production call is a 3-run consensus: median run reported, per-call kcal spread and item agreement disclosed - so any agent can probe reproducibility directly

Systematic errors we found (and corrected)

These are the classes of mistakes vision-language models made against scale-weighed truth - published as findings; the exact correction rules, anchor values and dish lexicons remain proprietary:

• Hidden staples: rice under donburi toppings and noodles under broth are systematically under-counted when occluded
• Fried-batter oil absorption: tempura and fried items read as their pre-frying calories; absorbed oil is invisible
• Cooked-vs-dry density confusion: pasta and rice portion estimates drift when the model anchors on dry-weight intuitions
• Western pan-sauce hidden fat: butter- and oil-based sauces carry substantial energy that has almost no visual signature
• Dish-name ambiguity: visually similar dishes with different energy profiles (e.g. grilled fish vs glazed chicken; Chinese dishes with shared plating) flip the estimate when misnamed
• Piece counting: models miscount uniform items (gyoza, karaage) under overlap
• Near-zero-kcal lookalikes: konnyaku and broth-heavy dishes read as denser food

portrait-analysis: 16-subject diverse-skin set

Personal color analysis is calibrated on a 16-subject set spanning diverse skin tones. Skin colorimetry is measured from pixels (CIELAB, ITA angle, cheek-region medians) and cross-checked against vision-model attributes; disagreement and photo-quality problems (overexposure, color cast) are flagged and downgrade the measurement to a reference value instead of silently pretending confidence. The 4-season verdict follows a published rule table with all season scores disclosed.

What we do not publish

The correction rule tables, numeric anchors, thresholds and dish lexicons are not published. The weighed ground-truth dataset and the regression harness that verifies every rules change are private. What you can always inspect: the per-call disclosures (spread, agreement, quality flags), free recorded samples at /samples/meal-vision, and the independent probe record at market402.com.

Scope and honesty notes

meal-vision is not for allergen detection or medical decisions. portrait-analysis performs no identification and images are not stored. Calibration numbers above refer to the rules version stamped in every API response (rules_version); the set grows over time and this page is updated when it does.

Cite this data

License: CC BY 4.0. Cite as: "Mart402 vision calibration - 30 weighed meals - https://mart402.com/vision-calibration". Worked agent flows: meal logging - personal color.