Integrating Licensed Odor Threshold Data into a Flavor Pairing Engine: Pair-Level Score Shifts and Unit Errors Caught

CompKitchen is a public flavor pairing engine that ranks ingredient combinations by shared volatilecompounds (food-science axis) blended with recipe co-occurrence (culinary axis). Prior to this work, anadvertised “perceptual weighting” feature — boosting compounds with low orthonasal odor thresholds inthe IDF-weighted Jaccard scorer — was effectively dormant for the ~5 weeks the public site had been live:the upstream literature-mining table was empty, so every compound’s perceptual weight defaulted to 1.0.Licensing the Rychlik, Schieberle & Grosch (1998) compilation of key food-odorant thresholds gave us adense, hand-curated source keyed to compounds already in our graph. Integration matched 191 of 224Rychlik odorants against a unified compounds_master record (147 by CAS auto-normalization, 44 by ahand-built CAS alias bridge for German spellings and stereochem variants), emitted 492 matrix-taggedthreshold rows, and seeded 432 perceptual weights for the live scoring engine on day one (the perceptualweight table has since grown to 1,772 compounds as subsequent integrations layered on top of Rychlik, butthe Rychlik-specific effect we report below is isolated to a clean before/after diff window taken inside theintegration interval). We report two complementary evaluations: a science-formula compound-axis subscore across 18 representative pairs (17 of 18 shifted, max +28.3% at apple + rose, driven by βdamascenone, orthonasal-water threshold ~0.0014 ppb), and a classic-formula full-pair-score evaluationacross 198 pairings spanning 20 test ingredients (all 198 shifted; median |Δ| = 0.000673, max |Δ| =0.004091 — a much smaller absolute effect, because the recipe co-occurrence axis dominates the classicformula). The integration also surfaced 39 compounds whose canonical orthonasal-water thresholdsshifted ≥2× in a clean before/after diff isolated to the Rychlik window, including a maple-furanone valuethat was 187,500× too low and a 1-octen-3-one value off by 613,700×. We report the integration pipeline,both evaluations, the corrections caught, and one open research question: 326 of Rychlik’s matrix-specificthreshold rows (oil, beer, starch, buttermilk, cellulose, meat) are stored in the schema but inert in scoring,pending a matrix-aware OAV extension.