Horizon started with a swipe. Two partners, one card, swipe right if you like it — and when both swipe right on the same flight or experience, you get a match. That mechanic worked in under a day. What took months was everything that came after the match: getting from "we both liked this" to "this is booked" without one partner silently backing out, re-searching, or abandoning the cart in a group chat argument.
Building Horizon taught us where group travel actually breaks. It is not at search. Couples do not abandon trips because they could not find a flight — they abandon because nobody agreed on which flight. Horizon was still architecturally a search app that happened to have a swipe layer on top. You queried inventory, scrolled results, matched, and then tried to force consensus after the fact. The swipe was the best part. The search-first foundation was the wrong frame.
New Horizon is not a refactor of Horizon. It is a ground-up redesign with a different objective: shift from Search to Alignment. The product question is no longer "what flights can we find?" It is "have both partners ratified this bundle before any inventory API executes?" That single inversion changes every architectural decision downstream.
What we kept from Horizon is the evidence, not the scaffolding. The swipe deck survived — partners still ratify or veto on a shared surface. Partner identity and Vault sync survived — budget, destination, dates, and travel vibe still filter what enters the deck. The Supabase backend and session-bound Duffel edge proxy survived — the client still never holds raw API keys. Affiliate checkout and disclosure UI survived — a ratified match still routes to fulfillment, not a dead-end list item.
What we designed new is everything Horizon was missing once we understood the real problem. The Consensus Engine owns match detection and bilateral ratification — a unilateral swipe is not enough. The Vault runs before search, not after — preferences constrain the query instead of filtering an infinite result set on the client. The Itinerary cache is offline-first — deck interaction does not trigger API round-trips. And governance is not an afterthought: OrchestrationManager, the Human Veto Queue, and the Triple Gate protocol (consensus → veto → checkout) mean every destructive action is blocked by default until a human explicitly confirms it.
Horizon also surfaced the need for an agentic layer — not chat for its own sake, but structured tool execution across the booking lifecycle. AgentBridge coordinates search, bundle assembly, and transit queue tools under the same governance constraints as the UI. The AI proposes. The Consensus Deck and Human Veto Queue dispose. Nothing executes silently.
Horizon did not fail. It succeeded as a prototype. It proved the swipe mechanic, stress-tested partner sync, and exposed every gap between discovery and booking. New Horizon is the Alignment Engine that Horizon pointed toward — built from those lessons, not patched on top of them. Horizon remains live as a demo of the original model. New Horizon is the flagship: align locally, ratify together, fulfill once, audit everything.
If you are building for groups, the lesson is simple. Search optimizes for one decision-maker. Alignment optimizes for consensus. Horizon showed us the mechanic. New Horizon is the architecture.