No retailer pays to rank higher.

When a query comes in, the matcher computes a hybrid_score per candidate product using four signals — three text-based (token overlap, trigram, brand similarity) and one semantic (embedding distance). The top-N candidates by hybrid_score become the result list, in score order.

hybrid_score =
  (emb_dist != null && trgm_sim > 0)
    ? (1.0 - emb_dist / 2.0) * 0.7 + trgm_sim * 0.3
  : (emb_dist != null)
    ? (1.0 - emb_dist / 2.0)
  : trgm_sim
  + token_overlap_pct * 0.30   // plurality-aware (Phase 12.22)

Within each matched product, offers are sorted cheapest total ascending where total_aud = price + shipping. The cheapest in-stock offer becomes best_offer. No retailer-specific weighting; no commission-aware re-rank; no “preferred partner” bypass.

// Materialise top-N offers in cheapest-first order. The first is best_offer.
const sortedOffers = [...offers]
  .filter((o) => o.in_stock && !isStale(o.last_seen_at))
  .sort((a, b) => a.total_aud - b.total_aud);

const bestOffer = sortedOffers[0];

When an agent calls the find_alternatives MCP tool, we evaluate candidate offers against a baseline product and return ranked alternatives. The ranking is governed by ADR-0007: a user-selectable trust slider that defaults to pure cheapest delivered total. The caller controls the trade-off; we never apply it silently.

• Default (trust_weight=0): sort by delivered total ascending. Two retailers with identical delivered totals tie — no thumb on the scale. This is the model that ships and the model every passive caller gets.
• Opt-in (trust_weight ∈ (0,1]): blend in a merchant trust_score (rating + delivery accuracy + dispute rate + tenure). At trust_weight=1 ranking is fully reliability-weighted; in between is a linear blend. The formula and the per-candidate inputs are returned in the response so the caller can audit the trade-off.
• Hard floors before ranking: every candidate must meet-or-exceed every declared baseline spec (IP67 ≠ IP69; we never silently downgrade). Undeclared candidate specs disqualify. Delivered total includes shipping + cross-border VAT differential (the Hungarian-27% / German-19% gap that bare list prices hide).
• List price only. Vouchers, promo codes, and loyalty pricing are neverread. They’re user-specific and irrelevant to apples-to-apples comparison.

// trust_weight = 0 (default) → pure-cheapest sort by deliveredTotal ascending.
// trust_weight > 0 → linear blend between price_score and trust_score.
const priceScore = 1 - (candidate.deliveredTotal - cheapestDeliveredTotal) / cheapestDeliveredTotal;
const trustScore = candidate.trustScore ?? 0.5;  // 0.5 = neutral when unknown
const weightedScore = (1 - trustWeight) * priceScore + trustWeight * trustScore;

Phase-14.7 disclosure: the trust columns (rating_score, delivery_accuracy_pct, dispute_rate_pct, account_tenure_start) are NULL by default in today’s catalogue. Until they populate, raising trust_weight above 0 reduces every candidate to the neutral default of 0.5 — i.e. it currently tie-breaks at random rather than meaningfully discriminating. We’re shipping the slider as API surface first; the consumer-facing UI on /search waits until the underlying data is real.

• The code is in plain sight. Apart from the SQL migration, every rank decision is in TypeScript at the paths above —apps/web/src/lib/search/run-search.ts for search, apps/web/src/lib/alternatives/ for the alternatives engine.
• The match_confidence field is exposed. Every response includes the four sub-scores (token_overlap, trgm, brand_sim, embedding distance) so an agent can audit why a result ranked where it did.
• The catalogue is read-only at rank time. Commission rates are stored in a separate table not joined into the RPC’s query plan. There is no syntactic path for commission % to influence ranking.
• find_alternatives is honest about what it skips. Every candidate that fails the meets-or-beats floor is returned in the skipped[]array with a per-candidate reason — spec downgrade, no shipping to buyer country, undeclared spec, list-price too high, etc. You see what didn’t qualify and why; we never silently drop honest losses to make the “cheapest” pick look better.
• trust_weight is the caller’s knob, not ours. Default is 0 (pure cheapest). When the caller raises it above 0, the disclosed formula and per-candidate trust_score breakdown ship in the response. No background re-ranking; no “experimental defaults”.
• The MCP wire shows everything. Agents see prices, shipping, VAT differentials, trust factors, and the sort order directly. If we cheated, the agent would surface the inconsistency.