Σ Statistics engine
How Trevo decides
Trevo is an autonomous growth engineer that does the work while you stay in command of what ships. Here is exactly how it reaches its conclusions, and the permissions that bound what it can do.
Always-valid statistics (mSPRT)
Trevo evaluates experiments with a mixture Sequential Probability Ratio Test (mSPRT), not a fixed-horizon p-value. Classic A/B tests are only valid if you decide the sample size up front and never peek. Every early glance inflates the false-positive rate. mSPRT produces always-valid p-values and confidence intervals: you can look at the result at any time, and the agent can act the moment significance is genuinely reached, without a peeking penalty.
Sample-ratio mismatch (SRM)
Before trusting any result, Trevo checks that traffic actually split the way it was configured. A sample-ratio mismatch, a chi-square test of observed vs. expected exposures, means something upstream is broken (caching, a redirect, a bot filter), and the numbers can't be trusted. An experiment that trips SRM is flagged and auto-paused rather than reported.
Guardrails and auto-pause
Every experiment runs behind safety brakes that check on each stats tick. Trevo auto-pauses a running experiment when it detects an SRM, a guardrail metric that degrades significantly (beyond a magnitude floor, not just a tiny blip), or a spike in SDK errors. A pause is reversible: it flips assignment back to control and alerts you; it never concludes the experiment or deletes data. Safety brakes run unconditionally, even when proactive autonomy is switched off.
What the agent may and may not do
The boundary is simple: the agent owns everything reversible; you own everything irreversible. It proposes experiments, writes variant code, runs them, measures results, and explains what happened. It only ever writes to branches. It opens pull requests for your review. It never merges a PR and never ships a variant. Merging and shipping are yours, always.
Autonomy you control
Autonomy is a simple per-workspace switch. When it's on, Trevo works between your visits, proposing experiments and opening pull requests for review. When it's off, it acts only when you ask. Either way it only ever prepares work; it never merges a pull request or ships a variant. An organization-wide kill switch halts all proactive autonomy instantly, and the safety auto-pause keeps running no matter what.
See what Trevo would test on your product.
Connect the repo and Trevo proposes experiments grounded in what it read. Or preview the ideas on your domain first — the scan is free and cites its sources.