# Reference meals and the next benchmark The nine original photo fixtures use visually estimated portion ranges. Reference calories combine those estimates with nutrition densities. Results therefore describe agreement with those references, not measured calorie truth. A shared yardstick makes the runs comparable, but uncertain references cannot establish a reliable real-world accuracy ranking from small differences. ## Source existing documented meals The next step is to find published meal photos or datasets that already include measured portion weights and documented calorie and macro references for the pictured food. Update, October 6, 2026: the Nutrition5k comparison completed on September 18. Read the findings: https://kcalsnap.com/journal/nutrition5k-calorie-benchmark/ The steps below record the reference-selection requirements. For each candidate source: 1. Check that the nutrition and portion weights belong to the food actually pictured. A recipe's total calories do not establish the amount on a plate. 2. Record the source URL, collection method, cooked or raw state, serving size, ingredients and nutrition provenance. Keep reported uncertainty visible. 3. Check licensing and permitted image use before copying or publishing assets. 4. Freeze the reference set before model calls. Keep reference values independent of the predictions being evaluated and distinguish new evaluation data from examples already used to develop the matcher. ## What to report Compare portion estimates, nutrition-record choices and whole-meal totals. Include missing components in plate-level results. Report USDA-matched and AI-estimated items separately alongside coverage, latency and cost. Different food groups may have different difficulty, so similar errors alone do not prove that those nutrition sources are interchangeable. Published data can have limitations and may have appeared in model training. Document those limits rather than claiming that a public dataset is unseen. Measured portions also do not make reference calories exact: recipe calculation, food composition and nutrition-label uncertainty still matter.