Community · E2 · artifact verified
Plan day trips with a typed checker
Jev Trip is an explainable day-trip planner where the LLM plans ahead and Jev chooses and checks: scope choices keep one day in one city, per-place choices rank every candidate with probabilities, and code owns routes, times and validation.
01 · Role in the system
What Jev does here
Every itinerary passes through typed gates before it reaches the user. Scope choices ask whether the request is genuinely one day, one city and planning-only - flagging multi-day, multi-city or booking asks instead of silently doing them - and each candidate place is judged by its own choice question with explicit criteria, landing on priority, alternative or needs_information with probabilities attached. Answers are zod-validated on the pass/flag/unknown schema, timeouts fail toward pass-with-flags rather than garbage, and the live demo shows the checks beside the plan. Deterministic code keeps everything factual: route times, opening windows, versioned itineraries - the LLM drafts, Jev decides what the draft may claim, and neither invents a fact.
Draft freely, claim only what passes the gates
The planner-verifier split is cleanly drawn: the LLM drafts the itinerary, and Jev decides what the draft may claim. Scope choices ask whether the request is genuinely one day, one city, planning-only — flagging multi-day or booking asks instead of silently doing them — and every candidate place is judged by its own choice question with explicit criteria, landing on priority, alternative, or needs_information with probabilities attached.
Facts stay deterministic: route times, opening windows, and versioned itineraries are computed by code, so neither the planner nor the judge invents a fact. The live demo shows the checks beside the plan, which turns the verification layer from a promise into something a user can watch working.
Graceful degradation as a feature
Answers are zod-validated on a pass, flag, or unknown schema, and timeouts fail toward pass-with-flags rather than garbage — the system's worst case is an itinerary with honest warnings, not a confidently wrong plan. That fail-toward-honesty posture is the record's most transferable property.
Scope limits are explicit: day trips only, route facts depend on the operator's configured services, and the project carries no external evaluation — 19 stars at review time, shared as a Discord show-and-tell. The pattern, however, generalizes beyond travel: any LLM-drafted artifact can be gated by typed scope and quality checks before it reaches a user.
02 · Control boundary
Where Jev sits
The planner-verifier split: an LLM drafts freely, Jev answers typed choice questions to enforce scope and rank candidates with probabilities, and deterministic code owns routes, clocks and validation - every decision inspectable in the UI.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Day-trip scope by design; multi-day or multi-city requests are flagged, not planned.
- Route facts come from the operator's configured services; coverage varies by region.
- Shared as a Discord show-and-tell project; 19 stars at review time, no external evaluation.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- liaoyuhua ↗Community · github · public · checked 2026-09-28