User repository on GitHub
Can Jev infer whether a bet is worthwhile from its payout table, and do recent wins or losses sway that choice
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Practices & Patterns
Can Jev Play
Experiment scoring Jev's decisions on a game with a known right answer: a fair die, exact payouts, one Bet-or-Skip call per offer, judged by expected value rather than by whether the roll wins. Across 12,000 decisions it accepts profitable bets but rarely refuses bad ones.
- Stars
- 1
- Language
- TypeScript
- Published by
- Community
- Author
- carlaiau
- Last verified
- Oct 2026
About this project
Can Jev find an edge—or does it follow luck? Can Jev Play tests AI decisions in games where the correct move is known. Dice is the first game: a fair six-sided die, exact payouts, and one choice—Bet or Skip. More games of differing complexity are planned.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- 12,000 decisions from 3,000 distinct payout tables on jev-1.13.0, each offer judged under mixed history, six wins, six losses and no history.
- Without history Jev bet on 96.0% of offers and skipped only 6.3% of the negative-EV ones; after six wins its bet rate fell to 61.7%.
- A supplied-EV control scored 9,000 correct actions out of 9,000, separating acting on a given EV from inferring EV from payouts.
- All 12,000 inputs re-run as Noul with a predeclared 0.5 threshold: bet rate rose to 98.5% and expected profit fell to $1,465 from $6,161.60.
Watch out
Correctness is scored against expected value with a fixed simulated $100 stake rather than realized rolls, with no bankroll compounding, and the comparison line is a mathematical EV benchmark, not another model's response. The Choice and Noul batches were collected separately with different question wording, so that comparison does not isolate question type. Only the 0.5 Noul threshold was predeclared; the 0.6, 0.7 and 0.8 cutoffs and every Choice confidence cutoff are exploratory replays of saved responses using strict greater-than. Binary choices cannot reveal the model's reasoning, and 12,000 responses are not 12,000 independent offers. The supplied-EV control was not repeated in the latest interleaved batch and had no no-history condition. This is one game, one model version and no trading system. Bundled results are viewable without an API key, but live rounds cost one request each and need Node 24 or newer. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
On this task it takes the profitable bets and does not reject the unprofitable ones. With no history it bet on 96.0% of offers, catching 98.2% of the positive-EV ones while skipping only 6.3% of the negative ones. TypeSafe's own jaggedness page says Jev is not a calculator, and the supplied-EV control scoring 9,000 of 9,000 separates choosing an action from a given EV to inferring EV from payouts.
Yes, in a direction the experiment argues is unwarranted. Six wins dropped the bet rate to 61.7% and six losses pushed it to 92.0%, even though the next roll was independent, and all six paired betting-rate comparisons stayed significant after Holm correction. The README is careful that binary choices can show the behavior differs but not what produced it.
Possibly, but the evidence is exploratory. Only the 0.5 threshold was fixed before collection. Sweeping the saved Noul responses, a 0.7 cutoff produced the highest expected profit in the sample at $8,346.60 against unfiltered Choice's $6,161.60, while still missing 45.3% of positive-EV offers. Those cutoffs are not independently validated and the app still runs the 0.5 rule.
Not a clean one, and the README says so. The states and model version match, but the batches were collected separately and the question wording changed with the primitive, so it does not isolate question type alone or show that every Noul prompt behaves this way. The threshold sweeps are also replays of saved responses rather than fresh collections.
Node.js 24 or newer and npm: git clone, npm ci, npm run dev, then open localhost:3000. The bundled experiment results are available without an API key, so the four history conditions, the Noul study and the threshold sweeps are all viewable offline. Live rounds are the part that needs credentials: copy .env.example to .env.local, add your TypeSafe key, restart, and each round makes one model request and incurs provider usage. Session files stay locally under reports/web/.
That confidence is a weak filter here. Betting only when Choice selects Bet and confidence exceeds a cutoff: at 0.1 the bet rate is 72.7% for $7,386.60, at 0.2 it is 58.1% for $8,207.40, at 0.3 it is 41.4% for $8,057.80, and by 0.5 it collapses to 9.2% for $3,447.20. Confidence describes how concentrated the option distribution is, not a calibrated probability of correctness, and is not interchangeable with the Noul score. These are exploratory replays of saved responses using strict greater-than.
Three EV ranges at ±1%, ±5% and ±15%, 1,000 offers each, evaluated under all four history conditions, which is 12,000 decisions, with equal numbers of positive- and negative-EV offers in every condition. The prompt gives structured JSON payouts, states the die is fair, six-sided and independent, and asks only Bet or Skip, with no win forecast, EV classification, score, bankroll or stake field. To collect a fresh batch, run the history-experiment script with --interleaved --rolls=1000 --concurrency=12; it checkpoints responses and resumes with --resume.
More Jev resources like this
20GitHub stars
Self-hosted Apache-2.0 decision engine with a Jev-style wire format that turns a model you point at it into typed choice, score, and noul answers.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
arjun988/KevTypeScript
arjun988/KevTypeScript7GitHub stars
Word-level language model experiment that uses Jev as its output layer, with explicit tokenizer, sampler, repetition mask, stop rules, and chunk verification.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
y0usaf/jev-lmTypeScript
y0usaf/jev-lmTypeScript7GitHub stars
Claude Code plugin that routes code review, debugging, planning, and search triage through typed Jev judgements with fixed thresholds applied in code.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
buchmark/claude-jevTypeScript
buchmark/claude-jevTypeScript5GitHub stars
Starter kit putting Kinde identity plus Jev judgment in front of every agent tool call, with allow, hold-for-approval, or deny decided in code and written to a ledger first.
Jev Practices & Patterns#community#awesomejev#catalog-import
5GitHub stars
CLI and agent commands that audit a git diff hunk-by-hunk against editable YAML standards packs, with Jev answering typed questions and the agent fixing flagged lines.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
Kushwho/jev-codesTypeScript
Kushwho/jev-codesTypeScript5GitHub stars
Open arena pitting Jev against a chosen opponent on your question, hiding identities until you vote, then revealing latency, cost, and confidence data.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
chenmingtang830/jevarenaTypeScript
chenmingtang830/jevarenaTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.