Detect
Catalume starts with a catalog snapshot, public commerce surfaces, and buyer intents that make sense for the store.
THE CATALUME LOOP
Watch the buyer. Change the signal. Prove the difference.
Catalume turns the AI shopping funnel into an observable loop: detect, diagnose, fix, verify, and attribute. The product shows what happened and how certain the evidence is.
Diagnose
Every non-win gets a named cause: invisible, misunderstood, weak comparison evidence, blocked, or another observed failure class.
Fix
The merchant sees a concrete, reviewable artifact — a fact, spec block, policy entry, schema patch, or content change — before approval.
Verify
The exact failed mission runs again, then neighboring intents and other AI legs are checked for collateral movement.
Attribute
Share of Recommendation trends and influenced-revenue estimates connect the result to a business signal when the merchant has the data.
REPLAY THEATER
The signature moment is the losing moment.
The replay makes the decision inspectable: buyer question → retrieval → comparison → rationale summary → cart and checkout walk.
buyerfind a giftable pour-over set
evidencespec table missing heat retention
outcomeplaced, not chosen
nextdraft a fix and re-run
FIDELITY BADGES
Strong evidence, clearly labeled.
Retrieval can be ground truth when a platform consumes the same catalog, MCP, or UCP substrate. Reasoning and checkout stages carry their own simulation boundary. Catalume never silently upgrades a modeled result into a consumer ranking.
GROUND TRUTH
Retrieved.The platform-facing commerce substrate returned this product or fact.
SIMULATED
Modeled.The official provider API or a labeled model stands in for the unavailable consumer surface.
WALKED
Checkout.The path is inspected without completing payment.
SEE THE SURFACES