The schedule that crosses things out and then lights up what she liked.

No. Every plan, re-plan, swap and toggle calls Qloo live; plans and re-plans also call the routing model. Captured Qloo responses are used only in unit tests.

Without keys, as in CI, the app refuses to plan rather than invent a week, and the end-to-end suite asserts exactly that. This page's replay is a recreation of three stored live plans; each step links to the real one.

No. The model picks where to look (the rung, the domain, the query) and writes one coaching line per card. Qloo picks the venue and the verifier checks it. A coaching line keyed to an ID no tool returned is dropped, and drinking language in coaching is dropped on every case.

No. It is clinician support only: no diagnosis, no efficacy claims, no patient logins. Every case in the app is fictional. Behavioral activation is the evidence-based step; Rekindle supports that step and is not itself evidence-based.

Taste entities and tags, a map point and constraint filters. The case name never leaves the server. The text typed into an intake love box is sent to Qloo to find matches, so the clinician types only the taste, and a small deny list refuses clinical words and contact details before any request (a screen, not a guarantee).

The week number, her energy level (1 to 3), her loves by name, the names of activities she rated as enjoyed (not the scores), and the open slots. Each tool call returns the placed card's Qloo ID, name, type, hour, distance and Qloo's explainability weights, or why the slot stayed empty. The case label and its map point never reach the model, though the places it is shown, with their distances, say roughly where she is. The care constraints (in recovery, quiet, wheelchair, budget, radius) are applied to every Qloo request on the server and are not written into the prompt. Because the model is never told whether she is in recovery, drinking language in a coaching line is dropped on every case, not only in recovery (lib/agent.ts, userPrompt and COACH_ALC).

A random 20 venues per arm, from plans for 20 fictional profiles, checked one by one against public listings by our build agent (Claude, an AI; not blind); inconclusive: 3 of the LLM's 20 and 6 of Rekindle's 20. Wrong means not found, closed at the planned hour, or serving alcohol to a patient in recovery.

Our automated checker over-flags (it matches by street address and misses large venues such as parks), so the manual audit leads. The bench run of 7 Oct 2026 predates 31 later plan-path changes (listed in DEMO.md).

The bench measured p50 17.43 s and p95 20.61 s from click; the coaching step dominates. It meets the 30 s gate we set, not the 15 s we aimed for. The week streams in card by card with the agent trace live, so the wait shows its work.

A coverage refusal. The sheet says how many outings it could verify and why each empty slot is empty; a wider radius is offered only with the clinician's consent. It never fills a slot with an unverified suggestion. Try Cal, the sparse-coverage case, on /judge.

Yes. New fictional case opens the intake: type 3–6 loves, confirm each against Qloo's live matches, pick a neighbourhood, energy, budget and constraints, and save. Week 1 then plans live.