Molecular Pairing Tool (MCP Modernist)

The MCP Modernist tool drives all molecular pairing analysis in this project. It is reliable, but several output behaviors are easy to misread and have produced wrong conclusions in the past. This page documents the correct workflow and the known traps. For the flavor combinations themselves, see Flavor Pairing.

Workflow

Verify IDs first (verbose)

Always resolve ingredient IDs with search_ingredients(query, verbose=true) before any pairing call. IDs are lowercased, sometimes compressed (darkchocolate, corn_grain for popcorn), and a wrong ID throws rather than guessing. Verbose also returns the compound count — your first signal of indexing depth.

Run pairings or dish analysis

find_pairings(id, verbose=true) for one-to-many; assemble the full ingredient set for a dish graph via generate_dish_png / create_share_link.

Read intersections only from verbose output

See the warning below. Non-verbose shared arrays are not pairwise intersections.

Critical caveat: verbose vs non-verbose

Caution

find_pairings with verbose=false returns a generic “characteristic compounds” list that is identical for every partner — it is NOT the true pairwise intersection. Only verbose=true returns the real shared-compound set per pair.

Example: non-verbose vanilla pairings return the same five compounds (2,3-butanedione, 2-phenylethanol, 2-phenylethyl acetate, (2E,6E)-farnesol, 3-(methylthio)propanal) for strawberry, almond, cinnamon, eggplant, and 20 others alike. Verbose banana and darkchocolate pairings, by contrast, return correctly varying intersections.

Only cite shared-compound claims from verbose output. Never build a bridge argument on non-verbose data.

Interpreting results

Similarity score

The sim score reflects whole-profile overlap (closer to a cosine over compound vectors), not the count of shared compounds. Two ingredients can share a single high-impact compound and still score low, or share many low-impact compounds and score moderate.

Sparse-indexing ingredients

Some ingredients have shallow compound lists and will systematically under-pair.

Ingredient	Compounds	Implication
Dark chocolate	13	Pairs almost entirely on Maillard/roasted compounds (pyrazines, furfural, guaiacol) + benzaldehyde + vanillin. Under-reads fruit affinities.
Banana	43	Broad ester/floral profile; strongest kin are other fruits.
Vanilla	41	Broad; functions as a connector across spice, fruit, and roasted families.

Note

Dark chocolate’s shallow index is why it consistently shows weak pairings with citrus and fruit. A weak sim between dark chocolate and a fruit is partly an artifact of indexing depth — corroborate with culinary reasoning before discarding a pairing.

Indirect bridges and keystone ingredients

A dish does not need every pair to overlap. The structurally important question is whether a keystone ingredient touches the otherwise-disconnected poles.

Worked example — dark chocolate + banana + vanilla:

Pair	Verbose overlap	Strength
Banana ↔ dark chocolate	benzaldehyde (+ phenylacetaldehyde)	Weak (1–2 compounds)
Banana ↔ vanilla	2-phenylethanol, farnesol, 2-phenylethyl acetate	Moderate
Vanilla ↔ dark chocolate	vanillin (+ guaiacol)	Keystone link

The two flavor poles barely connect directly; vanilla is the only ingredient touching both, so it is load-bearing, not decorative. When two target ingredients show a weak direct overlap, look for (or add) a keystone before concluding the pairing fails. This is exactly the structure behind the Dark Chocolate Banana Cake Gelato.

Technique-created bridges

Compound sets are not static across preparation. Browning, toasting, and caramelization generate Maillard compounds (furfural, pyrazines, guaiacol) that did not exist in the raw ingredient. A raw ingredient that under-pairs with a roasted one may pair strongly once caramelized. The tool indexes the named ingredient state; reason explicitly about what your process adds.

Parsing notes

Verbose find_pairings and search_ingredients in current sessions return native JSON objects directly (no text wrapper). If a call instead returns a text-wrapped payload (raw[0]['text']), json.loads the inner text. Confirm the shape per tool rather than assuming a single pattern.

---

Methodology reference. Update when tool behavior changes.