A pile of rows plus any yes/no questions you write. Each row gets a 0 to 1 score per question (0 = no, 1 = yes), all in seconds, for pennies. You choose the cut line; your rules apply to the whole dataset.
Presets are scored already, so switching is instant. Synthetic data from datasets.js; scores precomputed through the live function.
Gemini flash-lite writes the rows (synthetic, no real people); if it fails the built-in template generator is used instead. One generation per minute.
Jev (TypeSafe) is a model that does not write text. It answers typed questions about any text with a calibrated 0 to 1 confidence, about a quarter of a second and about $0.00002 per row. That makes it practical to run policy questions ("is this fraud?", "would this leak data?") across every row of a dataset or every action of an agent, and let code, not the model, decide what happens at each threshold. The key for Jev lives only in the serverless function; nothing you paste is stored. Sources: TypeSafe API · Confidence · jev-1.13 on OpenRouter · gating agent tool calls · independent evaluation.
Start with a pile of rows: 300 bank cases here. Paste your own or generate some later.