Fork-Cast (吃什么)

Deciding what to eat. The README argues traditional tools force you to label dozens of dishes by cuisine, spice, and price, or to scroll delivery apps for half an hour, while chatbots chatter for minutes without committing. Fork-Cast takes one plain sentence and returns a decision.

It is built on System 1 probabilistic decision models and the README says it fully supports TypeSafe Jev and Liquid AI D1. Instead of generating text, it computes probabilities and confidence, so the answer is a committed choice rather than prose to interpret.

Decisive pick: one candidate clearly wins, with confidence and a backup suggestion. Dilemma duel: two close candidates go to a single-spin wheel, no infinite re-rolls. Impasse: contradictory requests or an empty pool get an honest refusal. Blind box: a shrug like either is fine triggers an equal-probability draw.

Recently eaten dishes are down-weighted over a configurable window: 24 hours, 48 hours (the default), 72 hours, or one week. If every candidate was eaten inside the window, the system smooths the penalties so you never end up with no options.

Each device gets 10 free model decisions per day, resetting at midnight Beijing time. For heavier use you can paste your own TypeSafe or Liquid AI key in settings for unlimited private decisions; keys are stored encrypted in the browser and passed through only at the decision moment, never persisted server-side.

The README says your dish list, dining history, and preferences live in browser local storage, with no cloud user profiles, and BYOK keys are not written to the server. An invalid key produces an explicit message rather than silently spending the public quota.

A built-in mock simulation engine keeps the pick, duel, and anti-repetition flows working with no network or key. The project is published by the haoxiangxia GitHub account under CC BY-NC 4.0, which permits personal, research, and non-commercial use but forbids commercial resale, hosting, or repackaging.