An AI that answers with numbers, not paragraphs.
Jev does not write. You hand it a piece of text and a set of questions, and it returns a typed answer for each one with a probability attached. That makes its uncertainty visible, which is exactly what we want students and faculty to learn to read.
Yes/No
A yes or no question. Jev returns a single number between 0 and 1: the probability the answer is yes. 0.98 is a confident yes. 0.51 is a shrug.
Pick one
You define the options and describe each one. Jev picks one and shows the full probability spread, so you can see the runner-up and how close it was.
Rubric score
You write rubric levels in order. Jev returns a weighted score across them plus the probability of each level. A 3.9 of 4 means it is nearly certain of the top level.
Watch one request move through the graph
You are here. Run the example above and read the probabilities as they land.
Test five student coaches and two faculty tools
Switch between Student view and Faculty view in the header to see how the same numbers become different interfaces.
Build your own judgment graph
Pick a text, write questions, run it live, export the request as JSON to reuse in your own tools.
Judge live public data
Guardian headlines, BLS occupations, federal public comments, and 1900s newspapers, each run through Jev and charted.
Write your own coach
Your rubric, your checks, your feedback lines. Preview on Jev, save it beside the built-in coaches, share a link with students.
What Jev is, and what it is not
A generative model such as ChatGPT or Claude produces text: you ask, it writes. Jev does the opposite. You give it a piece of text and a set of questions you have written, and it returns an answer to each question as a number: a probability for a yes or no question, a probability for each option in a pick-one question, or a probability for each level of a rubric you wrote. It never writes a sentence, never explains itself, and cannot be asked to. That makes it useful for a narrow but real set of classroom jobs: giving a student a fast second reading of a draft against a rubric, sorting a stack of short responses so you know where to spend your attention, or turning a pile of real-world text into a chart students can argue with. Because the rubric is yours and the numbers are visible, students can see exactly what was asked and how sure the model was, which is the opposite of the black box they usually meet.
Issues to plan for: a vague question gets a confident wrong answer, so rubric wording matters more than with a chatbot. The model reads only the text in front of it and cannot check a source, know the assignment, or notice sarcasm. Its numbers look precise, and students will treat 0.97 as truth unless you teach them not to. The text you evaluate travels to a company's server for the length of the request, so it should not carry names or grades, and the page states that. And any triage tool can quietly become an assessment tool. The safeguard on this site is that every sentence of feedback is written by rules a person wrote, the model only supplies the numbers, and both are shown.
What Jev is good for
- Rubric-style judgments where you can name the levels.
- Sorting or flagging a batch of short texts fast and cheaply.
- Teaching probability literacy: the model shows its doubt instead of hiding it in fluent prose.
- Formative coaching, where a student revises and re-runs and watches the levels move.
What it is not
- Not a writer, tutor, or chatbot. It never produces text. Every sentence of feedback on this site is written by rules a human wrote.
- Not a grader. Treat every number as a prompt for human judgment, never as a verdict.
- Not an AI detector. It answers the question you wrote, and a badly written question gets a confident wrong answer.
- Not private by default: the text you evaluate travels to TypeSafe's servers for the duration of the request. Nothing here is stored.
Coaches and faculty tools
Student view shows only the coach's note. Faculty view adds every score, probability, and rubric behind it.
Pick a coach below. Students paste their own draft and choose Get feedback. They revise once using the Work on next list and run again to see what moved. Open the Teaching guide under any coach for timing and discussion prompts.
Build your own judgment graph
Paste any text on the left, add questions in the middle, run on the right. The JSON panel is the exact request your own code would send.
Coach Builder for faculty
Write your own rubric, your own checks, and your own feedback lines. Jev supplies the numbers; your words become the coach's note. Preview it live on the right, save it next to the built-in coaches, and share a link students can open.
Live data, judged
Each lens pulls real text from a public API, sends every item through Jev, and draws the answers. Confidence is always visible: faded marks are the ones the model was unsure about. Click any mark or item to see the raw answers.
Each lens takes under a minute. Switch chart types above the graph, hover any mark for the item behind it, and click to read the model's full answer. Faculty view adds the questions Jev was asked, which you can rewrite in Create.
Rate limits: the Jev proxy allows 60 requests a minute, so 40 items take under a minute. Regulations.gov's shared demo key allows about 10 fetches an hour; a free personal key from api.data.gov removes that limit and can be pasted in Connection. The Guardian developer key allows about 500 searches a day.