Asks a yes/no question and returns the probability that the answer is yes
, from 0 (no) to 1 (yes).
Returns noul (0–1). No separate confidence: two outcomes are fully described by one number.
Enter a public GitHub repo to get started.
Try
| PR# | Title | Type | Chance it breaks | Impact if it breaks | Main risk |
Reach | Explained | Test evidence |
Look first |
Reviewer | Advice | Jev cost |
|---|
Each cell shows Jev’s most likely answer and its probability. Hover a cell for the probability of every option, a column heading for the exact question, or a cost for its breakdown. Jev also sees computed facts: file types changed, areas touched, labels or branch, and flags like CI or dependency changes. What each column means.
How it worksHow Jev answers, how much it can read, what each column is, and a real Jev call
How Jev answers
Jev is a judgment model, not a text generator. It reads a state (here: computed PR facts, the title, the description and the diff) and answers typed questions in a single call. Each question’s type fixes the shape of its answer. TypeSafe’s docs define three. The examples on each card are real answers from the response below.
TypeSafe borrows the name “System One” from the contrast between fast, intuitive System 1 judgment and slower, deliberate System 2 reasoning. Jev is built for the first kind: quick, inexpensive judgments over structured state. Here it screens every PR with narrow questions and probabilities, leaving deeper investigation to reviewers when the answers are uncertain or risky.
Selects one option from a set that has no order between them
, and returns the most probable option.
Returns choice, probabilities for every option (summing to 1), and confidence: high when one option peaks, low when it’s spread flat.
Rates the state against ordered, descriptive levels
, for when the answer is a position on a spectrum.
Returns score, probabilities per level, a legend and confidence. The score is an average: each level number times its probability, added up.
How much Jev can read
32ktokens for the
state plus the single longest question64ktokens per request: the
state plus all 14 questions together~30kour budget for the PR text: 32k minus the longest question, measured, and a safety margin
Those are Jev’s limits per TypeSafe’s Models page. Jev reads the state once and answers every question against it. When a PR’s text fits the budget, Jev reads it whole. When it doesn’t, the worker trims the diff hunk by hunk rather than dropping whole files. Every changed file keeps its header, so Jev still sees the full shape of the change. Hunks are then added one per file per round, source first, then migrations, CI and dependencies, then tests, config and docs, with lockfiles, build output and binaries last, so one huge file can’t crowd out the rest. Each file notes how many hunks were left out, and because missing diff can change the answers, the row gets a ! Trimmed badge saying how much Jev read, with a matching ! beside its advice. A long description is shortened to about an eighth of the budget first. Putting behavior first also helps accuracy: TypeSafe notes Jev suffers from context rot, so unrelated material in the state costs you accuracy.
The budget assumes 2.75 characters per token: pessimistic for prose, but close for a diff of mostly file paths and hashes, so a PR near the limit can still be rejected.
What each column is
| Column | Type | Levels (low to high) or options | Question Jev answers |
|---|
- Every cell shows the most likely level and Jev’s probability for it. Hover a cell for every level’s probability.
- Choice where there’s no order. Type, Main risk, Look first and Reviewer are categories, so they’re Choice questions. Everything with a natural order is a Score, per TypeSafe’s guidance:
If the answer is a position on a spectrum, use a Score.
- Evidence flags are Nouls. Four yes/no questions about what the diff itself shows: weakened error handling, unsafe input handling, a broken contract (an API, flag, config key or format others rely on), and leftover debug code. Each appears as a badge under the title when Jev puts it at 50% or more; hover a badge for the question. Any flag at 60% or more shifts the Advice color one step toward red.
- Commits get the same columns. Pick Commits and a branch (the repo’s default to start) to assess what has already landed. A commit’s first message line is its title and the rest its description. Its facts cover the branch, merge commits, bot authors and co-authors instead of labels and forks. Because a commit is already on the branch, its questions ask what to do now, so Advice reads Leave it, Fix forward or Back it out.
- Jev cost isn’t one of Jev’s answers. We calculate it from the input tokens Jev reports, at Workers AI’s price plus Cloudflare’s 5% credit fee.