Autonomous, not unsupervised.

Approval

Nothing reaches your users without sign-off.

Every variant Curious generates lands in a review queue before it serves. You approve it or reject it, from the dashboard or wherever you work. Approval is on by default and stays on until you decide a surface has earned autonomy. Turning it off is one setting per experiment, and turning it back on is the same setting.

Guardrails

Hard limits, enforced with discipline.

You declare what can't break: a cap on unsubscribes, a floor on session completion, a refund rate that can't rise, phrases that must never appear and ones that always must. Content rules are checked before a variant ever enters review. Metric guardrails are enforced on live performance, and a variant that breaches one is pulled automatically.

Enforcement is statistically honest in both directions. Three support tickets in thirty decisions can read as 10% by pure chance, so noise like that never kills a variant. When one is pulled, the reason is stated in your terms: unsubscribe rate 2.9% exceeded your 2% cap across 212 decisions. And your baseline is never touched. Curious deactivates its own variants, not your control.

Visibility

See everything, including the reasoning.

Every variant carries the hypothesis it was built to test. Every conclusion explains what happened, for whom (this version performed 3x better on mobile than desktop), and, when the outcome didn't match the prediction, where the reasoning went wrong. Over time this becomes a ledger of what the system believes about your customers, with the evidence behind each belief. Beliefs the data refuted stay visible with their refutation, because knowing what didn't work is half of what you're paying for. Your experiments end in answers you own, not weights in someone else's black box.

You can read all of it from the dashboard, or from your own tools: the API and the MCP server expose the same picture to your agents and LLM clients, and your team can submit its own variant ideas from either.

Discipline

Statistics that never cheat.

The model only uses information that existed at the moment of each decision, never anything that arrived after. Its predictions are scored continuously against what production actually served, drift is monitored, and when something looks off, a human is warned rather than the system silently changing course. And model judgment is structurally capped: AI opinion can never outvote your data.

Brand

On-brand by construction.

At onboarding, Curious researches your business: your brand voice, your competitive landscape, and your customer personas, down to their jobs to be done, objections, and triggers. Generation happens inside that understanding plus the content rules you declare, and your team reviews the early rounds until the voice is right. The result reads like you wrote it, because the system learned from what you wrote.