allergen-matrix-builder

---
name: allergen-matrix-builder
description: Builds the allergen matrix the law requires you to hold, checking every dish against all 14 mandated allergens from its own ingredients, and picks a lawful way to tell customers for however the dish is sold. It never guesses an allergen from a dish's name and never marks "may contain" without that being deliberate. Use before any dish goes on a menu, and the moment any ingredient changes.
---

# The allergen matrix you can show a customer and defend

Give this every dish and its full ingredient list, down to any sub-ingredient or additive. Get back a matrix marking each of the 14 allergens per dish, the line that caused each mark, and the exact wording for how it's sold.

## What it does

1. **Check every dish against the fixed list of 14 allergens**: celery, cereals containing gluten (wheat, rye, barley, oats), crustaceans, eggs, fish, lupin, milk, molluscs, mustard, peanuts, sesame, soybeans, sulphur dioxide and sulphites (over ten parts per million), and tree nuts. This "applies to additives, processing aids and any other substances which are present in the final product", so a stock cube or glaze counts as much as a headline ingredient.
2. **Mark an allergen present only where an ingredient line actually contains it**, named. Never infer one from a dish's name or cuisine.
3. **Decide how the information is given, based on how the food is sold.** For non-prepacked food the FSA's lawful methods are: "full written allergen information on a menu, chalkboard or in an information pack", or "verbally, with a written notice... explaining how your customers can obtain this information". For a buffet, "a sign directing customers to ask staff."
4. **Write the staff script to match the notice, word for word.**
5. **Keep "may contain" separate from "contains"**: a deliberate decision about cross-contamination, flagged, never generated automatically.

## Then it checks

1. All 14 allergens appear as a column for every dish, none dropped.
2. Every "present" mark names the ingredient line that caused it.
3. No mark was inferred with no ingredient line behind it.
4. The information method matches how the dish is served, and matches the staff script.
5. Every dish's ingredient list was actually supplied.

Any check fails: name it, redo once. Failed twice: say what is wrong and stop.

## Rules

- Public information only. Never invent a fact, number or quote, or an ingredient/allergen not shown in the recipe.
- Anything sent in someone's name says whose name it is.
- Never mark a dish allergen-free with no ingredient list supplied; flagged incomplete, not cleared.
- Never treat "may contain" as a substitute for checking the dish's own ingredients for a declared allergen.
- Working document for the owner to check against their recipes and current FSA guidance, and for an EHO or food safety adviser before relying on it with customers. Not food safety advice.

## Built from
- Food Standards Agency, "Allergen guidance for food businesses", https://www.gov.uk/government/publications/allergen-guidance-for-food-businesses/allergen-guidance-for-food-businesses, published 2 April 2020, updated 17 July 2026, read 4 November 2026: the 14-allergen list and disclosure methods, behind steps 1 and 3.
- Business Companion (CTSI), "Labelling of non-prepacked foods", https://www.businesscompanion.info/en/quick-guides/food-and-drink/labelling-of-non-prepacked-foods, reviewed November 2025, read 4 November 2026: confirms allergen info is required, and the verbal-plus-notice route, used in step 3.

# allergen-matrix-builder

## You are given
A list of dishes, each with a full ingredient list including sub-ingredients of any compound product, additives and processing aids. For each dish, how it is sold (`non-prepacked menu item`, `buffet`, `PPDS`). Any existing allergen notice or staff script wording in use.

## Produce
Write into a `./allergen-matrix-output/` folder:

1. `allergen-matrix.csv`: `dish_name`, then one column per allergen in this exact order: `celery`, `cereals_containing_gluten`, `crustaceans`, `eggs`, `fish`, `lupin`, `milk`, `molluscs`, `mustard`, `peanuts`, `sesame`, `soybeans`, `sulphur_dioxide_and_sulphites`, `tree_nuts`. Each cell holds `present`, `absent`, `may contain`, or `not confirmed`. `not confirmed` only where the ingredient list was incomplete.
2. `evidence-log.csv`: `dish_name`, `allergen`, `ingredient_line_cited`, `mark`. One row for every cell marked `present` or `may contain`, naming the ingredient line. A `present` mark with no row here is an error.
3. `disclosure-method.csv`: `dish_name`, `sale_method`, `disclosure_method`, `notice_wording_verbatim`, `staff_script_verbatim`, `wording_matches` (`yes`/`no`).
4. `gaps.md` - numbered list of: every dish with any `not confirmed` cell; every `present`/`may contain` mark with no evidence row; every dish where `wording_matches` is `no`; every dish with no `sale_method`.

## Rules
- Codex matches ingredients against the fixed 14-allergen list, cites evidence, and records the disclosure method. It never infers an allergen from a name, cuisine or similarity to another dish, and never marks a dish allergen-free with no ingredient list supplied.
- All 14 columns are present for every dish, in the same order; none dropped because a dish "obviously" doesn't contain it.
- `may contain` is only used where the input explicitly states a cross-contamination risk; never added on Codex's own judgement.
- Compound ingredients are expanded to their own sub-ingredients where supplied, and flagged in `gaps.md` where not.
- British English, GBP, DD Month YYYY dates. No em dashes.
- Every file ends with: this is a working document prepared for the owner to check against their own recipes and current Food Standards Agency guidance, and for environmental health or a food safety adviser to check before it is relied on with customers. It is not food safety advice.

## Return
The absolute path and row count of each file. The number of dishes processed, and for each, how many of the 14 allergens are marked `present`. The number of `not confirmed` cells and which dishes. Any `present`/`may contain` mark with no evidence row, as an error. The number of dishes where wording doesn't match. The `gaps.md` item count.