Keyprint Research outputs · fixed public benchmarks
Same task. Two ways to choose the words. Inspect all 144 ordinary and marked pairs from GSM8K, SQuAD and IFEval. These are saved outputs from the current research SDK. Every selected task stays in the record.
The automatic score checks answers or instruction compliance. A separately blinded assistant review checks meaning and task adequacy. Neither represents human reader judgments.
Compare answers · Try your own text in the interactive demo
Choose a task Previous Next
Task and reference Reference used for scoring
Wording highlights show changed text, not changed meaning. Identical outputs are legitimate observations. A task passes jointly only when it ends normally, passes its strict automatic check, and passes its independent assistant meaning review.
One model, key and frozen sampling configuration. The 5-percentage-point primary margin concerns an equally weighted mixture of three fixed source/seed banks. Public training exposure is unknown; this is not an all-domain result or a Claude detector. Earlier failures remain part of the research record.
Some public benchmark references are wrong or ambiguous. Original references and all selected tasks are retained. Joint-score agreement alone does not establish semantic quality. Sources: GSM8K (MIT), SQuAD (CC BY-SA 4.0), IFEval (Apache 2.0). See pinned source hashes and revisions and dataset, evaluator and license notes . Original data retains its upstream licenses. SQuAD sampling uses one predetermined question per distinct passage. IFEval key 2785 was excluded before sampling because its prompt and checker disagree.