Community · E2 · artifact verified

Find city permits from a free-text plan

Describe what you want to do in San Francisco and a rules engine works out which permits you need - Jev answers each rules question as a typed choice, falls back to asking you when unknown, and summarizes fees and deadlines per permit.

01 · Role in the system

What Jev does here

The navigator drives a permit rules engine from free text, and Jev sits between the text and the rules: the engine's facts each become a Jev choice question over that fact's options plus an explicit unknown, and the most probable label is taken as the answer. When Jev says unknown - or the call fails - the question goes to the user instead, so the system degrades to a conversation rather than guessing. The first question gates everything: if Jev judges the plan is not an event at all, the flow stops early. Answers carry who supplied them (jev or user), the frontend surfaces the provenance, and the output is a per-permit summary with fees and deadlines.

A model in front of a rules engine, with an exit to the user

The navigator's shape solves the two failure modes of rules-driven civic tools: free text in, rigid questions out. Jev sits between the user's plan and the rules engine, answering each engine fact as a typed choice over that fact's known options plus an explicit unknown — and when the answer is unknown or the call fails, the question goes to the user, so the system degrades to a conversation instead of guessing.

The gating question shows the design's discipline: if Jev judges the plan is not an event at all, the flow stops early rather than walking the permit graph on a false premise. Every answer carries its provenance — jev or user — and the frontend shows it, so the user knows which facts were judged and which were told.

One city, honestly scoped

The rules engine encodes exactly one city's permit graph — San Francisco — and the record does not pretend otherwise; calls route through an OpenRouter-based client rather than the native endpoint, and the project is young with rules coverage described as the author's snapshot of city permitting.

What the record demonstrates for the legal cluster is the fact-extraction pattern with a humane failure path: typed choices over enumerated options, an explicit unknown that routes to a human, and provenance surfaced in the UI. Swap the rules engine and the pattern applies to any jurisdiction or compliance domain.

02 · Control boundary

Where Jev sits

Jev as the fact-answerer in front of a deterministic rules engine: each rule question becomes a typed choice over known options plus unknown, failures and unknowns fall back to the user, and provenance of every answer is tracked 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

  • San Francisco only; the rules engine encodes one city's permit graph.
  • Jev calls route through an OpenRouter-based client, not the native endpoint.
  • 0 stars; rules coverage is the author's snapshot of city permitting.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

  • kmangal ↗Community · github · public · checked 2026-10-03