Catalog Calibration Gate

How we measure whether your promo catalog is ready for category-level routing

4.1The principle

Every Kornerway deployment runs a one-time pre-flight check on the partner's promo catalog. The engine extracts 10-dimension taste vectors from each SKU, then computes pairwise cosine similarity between category-mean vectors.

If categories are distinguishable in taste space, category-level routing is production-ready. If they cluster too closely, the catalog needs richer descriptions before category routing goes live.

4.2The threshold

this dataset
ROUTING VIABLE
MARGINAL
ENRICHMENT REQUIRED
0.0 0.95 0.97 1.0

4.3This dataset's measurement

birthday dining events membership racing
birthday 1.0000 0.9804 0.9819 0.9931 0.9765
dining 0.9804 1.0000 0.9980 0.9914 0.9959
events 0.9819 0.9980 1.0000 0.9907 0.9980
membership 0.9931 0.9914 0.9907 1.0000 0.9855
racing 0.9765 0.9959 0.9980 0.9855 1.0000

This catalog: 20-40 character SKU titles only → all pairs in red band → enrichment required.

4.4The fix path

  1. Partner provides full SKU descriptions (200-500 words per promo, marketing copy depth)
  2. Re-run extraction pipeline — zero architecture changes, no model retraining
  3. Re-run calibration gate to verify cosines drop below 0.95

This is a quantified onboarding step, not a hidden limitation. Every Kornerway pilot starts here.

Category-Mean Dimension Vectors

Category LUX NOV AES SOC AUTH SVC SENS PLAN VAL WELL
birthday 0.3724 0.2108 0.3246 0.3092 0.1980 0.3641 0.3027 0.3953 0.3045 0.2179
dining 0.3597 0.2775 0.3458 0.3129 0.2807 0.3469 0.3550 0.3713 0.1594 0.1947
events 0.3497 0.2833 0.3171 0.3545 0.2672 0.3333 0.3573 0.3849 0.1774 0.1791
membership 0.3765 0.2799 0.3318 0.2841 0.2306 0.3390 0.2799 0.4063 0.2335 0.2080
racing 0.3759 0.2710 0.3224 0.3632 0.2801 0.3471 0.3718 0.3870 0.1649 0.1299

Extraction prompts and propagation parameters not shown — see internal IP disclosure protocol.

→ See Suppression