Open /simulations/iq-test to give Jev 24 original multiple-choice reasoning questions: eight numerical, eight logical, and eight text-based patterns. Each question states its assumptions and has a checked answer with an explanation. A complete run produces a numerical Estimated IQ, labeled Uncalibrated heuristic, alongside its raw score and category breakdowns. This is an educational practice set, not a standardized or normed IQ test; the estimate does not establish human IQ, a population percentile or general model capability.
The set includes six warm-ups, twelve intermediate questions and six challenges. Difficulty labels are editorial, not measured Jev difficulty. The added questions use multistep operations and distractors based on plausible mistakes, such as reversing a transformation or averaging speeds without accounting for time. Each category has two correct answers in each option position.
The default Local demo always picks A. It makes no model requests and shows how the answer sheet works. Its 6/24 score and estimated value of 78 are scripted illustrations, not Jev's performance or an empirical random baseline.
Select Live Jev, then Run test, to use jev-latest with the configured key. The run sends at most 24 sequential requests. Each answered question stays visible for two seconds before the next question; the last answer has the same reading pause before the final report. Each request contains only that question's prompt, assumption, and four choices. Reference answers, explanations, other questions and previous results are not sent. Symbol patterns are text; Jev receives no images.
The question view follows the run automatically. A player shows the current question, selected answer and correctness, with a Stop button. Selecting an earlier question suspends following while Jev continues; check Follow current question to rejoin the current question. Live answers show the reported confidence, choice probabilities and request time. Confidence does not determine correctness: code compares the selected option to the reference key. Reveal Checked answer and explanation to inspect that key.
After all 24 answers, the final report shows the numerical IQ estimate, raw score and percentage, total correct/incorrect/unscored counts, and numerical, logical and pattern breakdowns. Following moves the view and keyboard focus to that report. If you are inspecting an earlier answer, your view stays on it; select Final result in the answer-sheet header to open the report. Selecting any question returns to its prompt. Open How this estimate is calculated to inspect the assumptions.
The pure conversion in lib/iq-estimate.ts, version assumed-reference-v2, uses:
estimated IQ = round(100 + 15 × (correct − 12) / 4)
The assumed raw-score mean of 12 correct out of 24 and standard deviation of 4 answers are provisional design choices. They have not been measured in a human sample or validated for Jev. Twelve is the midpoint of this small set; the assumed four-answer spread maps its endpoints three scale deviations from 100. Changing these assumptions changes the estimate. The displayed whole number is rounded, not a claim of measurement precision. There are now 25 possible raw scores and 25 distinct rounded estimates, with each answer changing the estimate by three or four points. The endpoints remain 55–145; more questions add score resolution without demonstrating better calibration or guaranteeing Jev will miss more answers.
| Correct | Estimated value |
|---|---|
| 0/24 | 55 |
| 6/24 | 78 |
| 12/24 | 100 |
| 18/24 | 123 |
| 24/24 | 145 |
The 100/15 scale follows a common standard-score convention described in Pearson's assessment primer, pages 3–4. That source also explains that standardized reference values come from normative data. It does not validate this playground's assumed 12/4 reference values, its questions, or comparisons between Jev and people. No percentile, confidence interval, intelligence classification or clinical interpretation is inferred.
The estimate uses only checked correctness across a fully answered test. Confidence, response speed and provider errors do not add points. Scores outside 0–24 and other question-set sizes are rejected. The original 12 questions and their answer keys are preserved; 12 questions are added in question-set version reasoning-v2. A different question set needs its own scoring-method review.
Every response must be a choice, select an offered option, report confidence in [0, 1], and include exactly the four offered probabilities, each in [0, 1], summing to 1 within 0.02. Malformed replies remain invalid and stop the run. Provider failures remain failed. Neither becomes a substituted answer or an incorrect guess.
Stop, a hidden tab, a key change or leaving the page aborts pending work. Reset discards the run, and late responses cannot repopulate it. There are no automatic retries or resumes; Run test starts a new full run. Questions after the stopping point remain Not run. Only 24 valid answers yield a final percentage; otherwise the page shows the answered count and correct subtotal.
The mode preference persists in browser storage. Results do not survive refresh. Export saves the question set, checked explanations, results, summary including category breakdowns, run status, requested model and execution mode under reasoning-v2. The iqEstimate object includes its numeric value, method version, calibration: "uncalibrated", assumed raw mean/deviation and output scale parameters. An incomplete test has iqEstimate: null and no overall percentage; an incomplete category has no percentage. Exports exclude credentials and raw provider errors. An exported report is an inspectable record, not proof of provider authenticity or an independently certified benchmark. Usage remains in the playground's shared Usage panel; missing or cancelled usage is not treated as free.
Uniform random choices have an expected score of 6/24 because each question has four options and one checked answer. This expectation describes repeated complete tests, not a measured result or a pass threshold.
tests/iq-test.test.ts independently works the reference solutions and checks request isolation, response validation, complete/partial scoring, the request cap, failures and cancellation. tests/e2e/iq-test.spec.ts covers local and mocked live runs, errors, reset, key changes, mode persistence, keyboard inspection, light/dark themes and narrow/short layouts. Automated tests make no live provider calls.