jev-for-chrome: finite actions on practice quiz pages

Viral “auto exam” demos aside, the maintainable path is: page state → finite Choice → deterministic clicks. Use simulation/practice pages only (ACP-style mock quizzes). Never for real certification cheating. Public-repo walkthrough, not a live re-run.

Source · chy4pro/jev-for-chrome · Open-source project

2026-09-22 deepened for practice-quiz automation; demos only — no real exam cheating. Source walkthrough, not a live re-run.

Problem

This walkthrough follows the public jev-for-chrome repository: observe a page, choose a bounded action, then let code execute it. Use only a practice page you own or may automate. Real certification exams and cheating are out of scope; JevLog has not reproduced the project.

Before you begin

• Stable Chrome; ability to load an unpacked extension (see repo README).
• A practice/simulation page you own or may automate — never a live proctored exam.

Sources

• GitHub repository

GitHub Open Graph preview for the chy4pro/jev-for-chrome repository; a repository preview, not a screenshot of running software. Notes are based on its public source.

Steps

1. Lock the scenario: practice page, not a real exam

Name the target: your local practice page, a public mock quiz, or the repository demo. Never point the automation at a real certification exam.

2. Install the Chrome extension or wire an ultrafast loop

Load the extension per its README, or run an observe–decide–act loop in a host. Begin read-only; minimize site permissions, avoid password fields, and keep auto-submit off.

3. Bound the next step as Choice: CLICK / TYPE / SELECT / DONE

Allow only listed actions such as CLICK, TYPE, SELECT, and DONE; never accept arbitrary JavaScript. Put page state and options in the request. On uncertainty, stop for a person.

from typesafe_sdk import Choice, TypeSafeClient

ACTIONS = {
    "CLICK": "Click a named control",
    "TYPE": "Type into an input",
    "SELECT": "Pick a radio/select option",
    "DONE": "Practice goal finished",
}

with TypeSafeClient() as client:
    res = client.system_one(
        state={
            "page": "mock quiz (not a real exam)",
            "stem": "Which option describes auto-scaling?",
            "options": ["A …", "B …", "C …", "D …"],
        },
        questions={
            "next": Choice(
                instructions="Pick exactly one next action; do not guess DONE.",
                criteria=ACTIONS,
            )
        },
    )
print(res.choices["next"].choice)

4. Deterministic code executes; the model never owns write APIs

Let deterministic code run the returned action, read the page again, and save a local trace. Require human confirmation for any submit action, and only on a practice page.

5. Acceptance proves “learned”, not “farmed”

Check the flow on three pages you control: multiple choice, fill-in, and end state. Confirm actions stay bounded, unknown cases stop, and traces can be replayed. Video speed is not acceptance evidence.

Check the result

You can run extension or loop → finite Choice → code execution → checkable traces on your own practice pages, and explain why real exam cheating is out of scope.

• Docs state practice/demo only; no live exam entry points.
• Action set is finite (CLICK/TYPE/SELECT/DONE or equivalent).
• unknown / low confidence halts; no auto final submit.

Continue reading

• jev-browser-use: next-step Choice on the DOM/a11y tree — not free clicking
• jev-ultrafast-mcp: a fast MCP judgment channel that keeps host permissions

FAQ

Can I use this on a real cloud certification exam?

No. Live exams forbid automation and cheating. This guide covers public mocks or pages you own.