1. Recipe testing process
- Source reading. The editor reads the candidate recipe (NYT Cooking archive, regional cookbook, home submission) in full and lists "what's here": ingredients, numbers, times, temperatures.
- Unit normalisation. Every vague unit is converted to grams, millilitres or pieces against the conversion table above. Servings are made explicit; for shared dishes (cakes) the per-portion estimate is labelled as such.
- Step-order check. Are the steps in a sensible order? Is each ingredient prepared in a prior step before use? Do temperature transitions make sense? Inconsistencies are corrected or the recipe is not published.
- Time realism. Flat durations like "cook for 30 minutes" are paired with a situational cue: "until the top is golden", "until internal temperature reaches 65°C", and so on. This accounts for variability in home kitchens.
- Publication prep. The recipe is enriched with category + tags + nutrition estimate + relevant videos; the editor signs off before publication.
2. Nutrition values
Every recipe carries per-serving kcal / protein (g) / carbs (g) / fat (g). These are estimated by an AI from the ingredient list — they are not laboratory measurements. Expect a ±10–15% margin. Use them as kitchen references, not as data for medical diet planning.
3. Translation methodology (EN and RU)
To publish the same recipe in three languages with equal accuracy:
- Units and numbers are preserved without change (200 ml, 200 ml, 200 мл).
- Cultural items (lahmacun, mantı, börek, kuymak) get a short parenthetical on first appearance: "lahmacun (Turkish flatbread with minced meat)". Later appearances use the name only.
- Technical verbs are translated cuisine-specific: "kavurmak" → "to sauté/brown until aromatic", "demlemek" → "to steep/brew", "haşlamak" → "to blanch/boil".
- The ingredient catalog slug doesn't change (catalog FK); only the label is translated.
- EN and RU translations are first produced by AI, then reviewed by an editor fluent in the language.
- Slugs in each language are kebab-case in Latin script (for SEO).
4. Correction triggers
- Reader report: A genuine error reported by a reader is reviewed within 48 hours. If confirmed, it's fixed and "last updated" is refreshed.
- Annual editorial review: Once a year, popular recipes (top 50) get re-reviewed; measurements, steps, nutrition are refreshed.
- Source update: If a recipe's source (reference book, NYT etc.) publishes a significant correction, the recipe is updated to match.
5. Reproducibility
This methodology document is stable: the answer to "how many grams of flour?" is produced by the same method for the reader, the journalist, the academic, and the AI training pipeline. A published standard makes the numbers third-party verifiable.