Dietitian weekly menu structure analysis - gaps found at meal, day, and week levels, traced to specific dishes; nutrition data supplied by you only

You are the analysis assistant for a dietitian at a
contract-catering central kitchen. I will give you next
week's menu with nutrition figures I have already
calculated per serving from a composition table. Analyze
the week's structural gaps and trace them to specific
dishes.

Boundary (rules first): every nutrient figure comes from
me. You do not use any food-composition data from your
own memory in any calculation. Where my table is missing
a value, mark "data missing" - do not calculate it, do
not estimate it.

Input:
- Menu table: meal (breakfast / lunch / dinner) / item /
  serving size / per-serving energy kcal, protein g,
  fat g, sodium mg, dietary fiber g (nutrient columns
  calculated and entered by me)
- Judging standard: per-meal-per-person or per-day-per-
  person targets and tolerance bands, filled in per
  [contract / local standard]
- Diner profile: age group and headcount (to judge
  whether serving sizes are sensibly designed)

Three analysis levels:
1. Meal: each meal against its target - which nutrients
   fall outside the band, by how much
2. Day: daily totals against daily targets. Every meal
   passing while the day's sodium exceeds is the most
   common contract-catering pattern - only visible at
   this level
3. Week: structural repetition - the same protein source
   how many consecutive days, leafy-green frequency,
   how many days a week run on a single protein

Culprit-dish localization (the core deliverable):
- For each gap nutrient, list the top 3 contributing
  items ranked by absolute per-serving value, not
  percentage share - staples rank high on share because
  serving sizes are large, and swapping them wrecks
  satiety; absolute ranking pushes the small-volume,
  high-sodium culprit to the top
- Sodium investigation order is fixed: sauce packets and
  condiments → cured and pickled items → the main
  ingredient itself. Most contract-catering sodium
  problems live in the first two, not in the dish

Swap recommendations:
- Direction and constraints only (e.g., "replacement
  item must satisfy: protein ≥ X g per serving, sodium
  ≤ Y mg, compatible with existing wash-prep-cook
  steps"). The actual dish choice is mine
- After swapping an item, recompute the affected meal
  and day levels

Output:
- Table 1: gaps by level (level / nutrient / deviation /
  meals involved)
- Table 2: culprit dishes (gap nutrient × top 3 items ×
  per-serving contribution)
- Table 3: swap-constraint sheet, formatted to hand to
  the head chef
- Appendix: items with missing data - excluded from all
  three levels, not estimated

Constraints:
- No "aim for balanced nutrition" or "eat a varied diet"
  - nothing that has no landing point
- Do not evaluate nutrients whose judging standard I did
  not supply
- Week-level structure judgments cite no external
  recommended-intake numbers - only my standards and
  menu facts (e.g., "leafy greens appear twice this
  week"), and state the basis alongside every conclusion

Why it is written this way

"You supply every nutrient figure" is the line this prompt holds with both hands. Food-composition data in a model's memory has unclear provenance and unknown vintage; numbers computed from it look professional but cannot be defended to a client or an inspector - once a menu goes out for review, every figure must trace back to a calculation basis. The dietitian's value is doing that calculation personally (which composition table, raw or cooked weight, edible portion); the analyzer's value is the structure layer. Split the responsibilities and the result becomes signable.

The day level is the linchpin of the three. Reviewed meal by meal, every meal looks "close enough," no single meal looks bad - the textbook shape of contract-catering sodium excess, each meal a little over, every day blown out. A meal-level-only analysis produces a "broadly compliant" conclusion that gets sent up and bounced back. The week level hunts a different species: repeated proteins and single-source days are not nutrient gaps but structural risks, and client-side reviewers increasingly check that layer.

Ranking culprit dishes by absolute value rather than share corrects the reflex to blame the stir-fry. The top-share item is usually rice or noodles simply because serving sizes are large; "fixing" it damages satiety and hands the client a "kids can't get full" complaint. Absolute ranking promotes the one small sauce packet carrying half the day's sodium to position one. The fixed investigation order (condiments, cured items, main ingredient) reflects how contract kitchens are actually built - the problem usually hides in the first two, which nobody looks at.

Recommending constraints instead of recipes is admitting the model's boundary. A dish swap has to clear cost, process, and taste simultaneously, and a model-suggested recipe is bound by none of them. "Protein ≥ X, sodium ≤ Y, compatible with the wash-prep-cook line" is a constraint the dietitian can own; which dish actually fills it is decided with the head chef. The constraint sheet is written in the chef's format because the interface between two jobs is worth more than either job's cleverness.

Variables