Skip to main content

Cookpit v3.2 — Source-Content Categorisation Rules

The chef-detective handles five distinct source-content categories consistently across every v3.2 file. This document publishes the categorisation rules so the AI Chef does not have to re-derive them per recipe.

Referenced by bundle/v3.2/prompt.md (in the deductive working order) and by bundle/v3.2/rules.md (as part of source-faithful handling under rules A2 / K1 / O).

1. Why this document exists​

Source recipes routinely carry non-method content — recipe tips, sponsored adverts, paywall chrome, source typos, make-ahead notes, and post-cook hints — that the chef-detective handles case-by-case. Each is handled consistently, but the categorisation rule was only in the chef-detective's head. This document publishes the rule.

The rule applies in stage C of the canonical fingerprint normalisation (filtering before tokenisation) and in the prompt's phase-1 resource-selection step (deciding what enters recipeInstructions[], what enters prereqs, and what is silently filtered).

2. The five categories​

Every piece of source content not in the explicit method body falls into one of five categories. The handling differs per category.

2.1 Sponsored content / brand placement​

Examples seen in the active corpus:

  • Try our app (BBC Good Food sources — carbonara, boeuf bourguignon)
  • BECOMEAMEMBER (Great British Chefs paywall — pork-fillet-braised-cheeks-and-pork-belly)

Handling:

  • Filter from recipeInstructions[] entirely. Sponsored content is not a culinary instruction; including it misleads schema.org consumers into treating advertising as part of the recipe.
  • Filter from the active-number sequence in stage C of fingerprint normalisation.
  • Document the exclusion in a prerequisites.notes[] item so the detective's filter is auditable.

Detection heuristic: Match the source line against the published v3.2.0 sponsored-content allowlist (see canonical-fingerprint-normalisation.md §5):

BECOMEAMEMBER
Try our app
You have <N> remaining read(s) today
Finish Ultimate Plus

Future revisions extend this list. Implementations SHOULD allow configuration with brand-specific patterns.

2.2 Paywall chrome / website navigation​

Examples seen in the active corpus:

  • You have two remaining reads today (Great British Chefs — pork-fillet-braised-cheeks-and-pork-belly)
  • Page-number/nav references that bleed into the extracted text
  • Timestamp metadata captured by PDF extraction in some sources

Handling: identical to sponsored content — filter from recipeInstructions[] and from the fingerprint sequence; document the exclusion in a prereq note.

Detection heuristic: lines that are clearly UI text rather than recipe content. The patterns vary per source platform. A productised detector should:

  • match against the published v3.2.0 chrome-pattern allowlist
  • accept user-configured allowlists for new sources
  • err on the side of preservation when ambiguous (false-positive filtering removes culinary content; false-negative filtering only adds harmless noise)

2.3 Culinary explanation tip​

Examples seen in the active corpus:

  • "The classic red wine to use in beef bourguignon is a burgundy (pinot noir), but any dry red that you would happily drink works." (boeuf)
  • "Beef bourguignon benefits from getting good-quality, well-marbled meat from the butcher's shop." (boeuf)

Handling:

  • Preserve verbatim in recipeInstructions[] as a 'Recipe tip: <text>' HowToStep. The Recipe tip: prefix distinguishes them from regular method steps for downstream consumers.
  • No structural encoding — these are advisory, not actionable.
  • Filter from the active-number sequence if they appear in a tips-block segment (per stage C of normalisation).

2.4 Structurally-actionable tip (make-ahead, leadTime,​

post-cook hint) Examples seen in the active corpus:

  • "Press the belly with the garlic, parsley and salt for 8 hours between two trays" (pork-fillet-braised-cheeks-and-pork-belly) → encoded as prereq.ingredients[].leadTime: "PT8H", and the 8-hour press itself is realised as a cookpit.prepCook phase

Handling:

  • Encode structurally wherever the schema offers a primitive (prereq.leadTime, the cookpit.preCook phase block, the cookpit.prepCook phase block).
  • Preserve in recipeInstructions[] as a 'Recipe tip: <text>' HowToStep so the source's own words are auditable.
  • Filter from the active-number sequence as tips-block content.

2.5 Source typo​

Examples seen in the active corpus:

  • "sweat of the carrots, onions and onions is a large saucepan" (pork-fillet-braised-cheeks-and-pork-belly, Step 1) — three errors: "of" should be omitted, "onions and onions" duplicates, "is" should be "in"

Handling:

  • Silently correct in tasks[].action — the chef-detective's job is to deduce CULINARY truth, not perpetuate source defects.
  • Preserve verbatim in recipeInstructions[] — the source's words are kept for source-faithful pass-through.
  • Document the correction in a prerequisites.notes[] item so the silent correction is auditable.
  • For numeric typos (none observed in the active corpus, but the pattern matters): the active-number sequence reflects what the source SAYS, not what the chef thinks the source MEANT. Numeric typos enter the fingerprint as written.

3. Decision tree​

Source content not in the explicit method body
│
├─ Is it advertising or brand placement?
│    YES → 2.1 sponsored content
│             [filter from recipeInstructions, fingerprint;
│              document in prereq notes]
│
├─ Is it website chrome / paywall / nav text?
│    YES → 2.2 paywall chrome
│             [filter; document]
│
├─ Is it a typo (semantic or factual)?
│    YES → 2.5 source typo
│             [correct in tasks[].action; preserve verbatim in
│              recipeInstructions[]; document in prereq note]
│
├─ Is it actionable as make-ahead / leadTime / post-cook?
│    YES → 2.4 structurally-actionable tip
│             [encode structurally; preserve verbatim in
│              recipeInstructions[]; filter from fingerprint]
│
└─ Otherwise (advisory culinary explanation, ratings, attribution etc.)
        → 2.3 culinary explanation tip
                [preserve verbatim in recipeInstructions[];
                 no structural encoding;
                 filter from fingerprint if in tips-block]

4. Worked example: pork-fillet-braised-cheeks-and-pork-belly​

The Stephen Crane source (pork_three_ways.pdf) contains:

Source contentCategoryHandling
BECOMEAMEMBER (×3, paywall)2.1 sponsoredFiltered from recipeInstructions; documented in prereq note q82e211144f
You have two remaining reads today (×3, paywall)2.2 chromeFiltered; documented
sweat of the carrots, onions and onions is a large saucepan (Step 1 typo)2.5 source typoCorrected to "carrots and onions in a large saucepan" in tasks[].action; verbatim in recipeInstructions[]; documented in prereq note q23389da5bd
Press the belly with the garlic, parsley and salt for 8 hours (Step 2)2.4 structurally-actionableEncoded as prereq.ingredients[].leadTime: "PT8H" (q15141f2498); verbatim in recipeInstructions[]; documented in prereq note q44554d5ffd
(no culinary explanation tips in the pork source)––

Each prereq note in the active corpus documents its own filter decision; this rule consolidates the categorisation so future files follow the same pattern without re-derivation.

5. Conformance​

A v3.2 file conforms to this rule when:

  1. Sponsored content and paywall chrome are absent from recipeInstructions[].
  2. Source typos are silently corrected in tasks[].action and preserved verbatim in recipeInstructions[].
  3. Each filter / correction is documented in a prerequisites.notes[] item.
  4. Structurally-actionable tips use the schema's existing primitives (prereq.leadTime, the cookpit.prepCook and cookpit.preCook phase blocks) rather than free-text prose.
  5. The cookpit-active-number-sequence-v3.2.0 fingerprint omits all five non-method categories.

The validator's V-LEX-FORBIDDEN and V-LEX-PERSONA-DRIFT checks catch the lexicon side; the categorisation itself is editorial and not (yet) machine-checkable beyond the published allowlists.