JF100 compares Jev 1.13.0 against Qwen3.5 0.8B, 2B and 4B with thinking disabled, a 512-token thinking budget, and a 2,048-token budget.

• 100 original English multiple-choice questions, 50 paired templates, 10 domains.
• 3,000 complete responses across ten published conditions; all valid final answers.
• Jev: 77.0%; 2B/2,048: 82.0%; 4B/2,048: 96.7%.
• 4B/2,048 exceeds Jev by 19.7 percentage points (paired-template 95% interval +12.7 to +26.7).
• The 2B/2,048 difference is +5.0 points (−2.3 to +12.3): descriptively close, not proven equivalent.

The release includes original questions, answers, exact-match scoring, compact per-request outcomes, offline recomputation, model/runtime fingerprints, and publication-ready charts. Full requests and responses, including thinking and token logprobs/top-5 alternatives, are provided in the compressed attachment. Check manifest.json for its SHA-256 and export transformations. Local filesystem paths in response routing metadata are replaced with model blob filenames; input messages, thinking, answers and LP values are unchanged.

This is an exploratory, post-observation release scope. It is not an IQ test, a universal intelligence upper bound, or a model parameter-count estimate. The AI-assisted dataset has not undergone independent human expert review. Confidence proxy definitions and limitations are documented in the repository.