Soupbase
Bilingual turtle-soup (lateral-thinking) puzzle site where Jev acts as host, judging player questions and grading solutions.
Overview
Soupbase (汤底) is a text-centred Chinese and English site for turtle-soup lateral-thinking puzzles, with Jev as the host. Players ask yes/no questions, get stepwise hints, submit reconstructions of the story, and can create private puzzles with revocable share links. Jev judges each player question with three host answers (yes, no, irrelevant) and grades submitted reconstructions, returning structured options, probabilities and confidence that the site displays.
When the model has no clear leaning (confidence below 0.45), the answer is shown as 'cannot confirm' along with the native confidence. Application code requires supported facts, a coherent explanation and sufficient confidence before marking a puzzle solved. It supports TypeSafe's API, OpenRouter and Vercel AI Gateway, either bring-your-own-key or a site key, with no automatic fallback between them, and lost responses are recovered by request ID rather than retried. It runs on Next.js with PGlite locally or Neon Postgres on Vercel. Code is MIT; puzzle content has separate licences.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Answer to the player's question about the puzzle | yes · no · irrelevant |
| Choice | How well does this submitted reconstruction match the true story?(inferred) | — |
How it works
- Player asks a question about the puzzle
- Jev answers yes, no or irrelevant with confidence
- Below 0.45 confidence, show 'cannot confirm'
- On a submitted solution, Jev grades it and code checks facts and confidence before clearing
Strengths
- Explicit uncertainty shown to players
- No silent retries or provider fallback
- Works without a key for browsing and creating
Caveats
- Model calls incur provider costs
- Chinese-first documentation
Getting started
npm ci && cp .env.example .env.local && npm run content:sync && npm run dev, then add a key in settings.At a glance
- Type
- App
- Status
- beta
- Language
- TypeScript
- Runtime
- Next.js web app (Vercel/Neon or local PGlite)
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Gaming: uses Jev `Choice` judgments to answer lateral-thinking puzzle questions and assess proposed solutions, with application code requiring supported facts, a coherent explanation, and sufficient confidence before marking a puzzle solved.
yibie/awesome-jevSoupbase is a bilingual Chinese-English Turtle Soup game where Jev judges player questions and reconstructions, and the app checks structured Choice results and confidence to decide clearance.
logicrw/awesome-jev-projectsJev `Choice` judgments answer lateral-thinking puzzle questions and assess proposed solutions, while application code requires supported facts, a coherent explanation, and sufficient confidence before marking a puzzle solved.
cobanov/awesome-jev
Frequently asked questions
What does Jev decide?
Whether each player question gets yes, no or irrelevant, and how well a submitted solution matches.
What happens when Jev is unsure?
Below 0.45 confidence the answer is shown as 'cannot confirm'.
Which providers work?
TypeSafe, OpenRouter and Vercel AI Gateway.
Related projects
Instrumented 2048 web lab where every move is a Jev Choice over four directions, with probabilities, latency and cost shown live.
ProjectbetaTypeScriptBrowser Minesweeper demo where Jev answers 'is this cell a mine?' for every frontier cell in one request, with an exact solver as fallback.
ProjectprototypeJavaScriptPandu: tiny local policy networks (2.9K-5K parameters) with calibrated uncertainty that fall back to a teacher when unsure.
ProjectresearchPythonPokémon Red on PyBoy where code owns route and arithmetic, Jev picks only at branches, and battle faint predictions are Brier-scored.
DemoprototypePythonTwenty-questions game where Jev's confidence decides when to stop: it commits, hedges or refuses to guess.
DemoprototypeTypeScriptReinforcement-learning lab where cached Jev judgments act as reward signals for DQN agents on four classic control and grid games.
ModelresearchPython