Observable behavior first. Interpretation second.

What a pilot does

1. You define one authorized preview or staging URL and a journey with a clear end state.
2. Six simulated sessions independently attempt it in real Chromium.
3. The run stores browser actions, screenshots, outcomes and model rationale.
4. Analysis separates observable behavior from model interpretation and recommendations.
5. After a fix, we rerun the same panel, goal, device and configuration.

Evidence labels

• Observed: directly present in the trace or target state—completion, repeated clicks, navigation, scrolling or missing content.
• Corroborated: independently observed across multiple sessions or verified against the page.
• Inferred: a plausible explanation that still needs human judgment.
• Recommended: a proposed change, not a measured outcome.

Comparison semantics

The panel parameters, goal, browser and run configuration are held constant. The underlying model is stochastic, so a click-by-click replay is not guaranteed. A baseline finding is called fixed only when the relevant page or state is exercised again and the behavior no longer reproduces. If the rerun never revisits that state, it remains unverified.

What JinuGen does not establish

• Population-level conversion, preference or statistical representativeness.
• Accessibility conformance or certification.
• That model rationale is a literal human thought process.
• That a clean run proves the absence of defects.
• That synthetic sessions replace functional testing or real customer research.

Model and interface boundaries

The browser system uses screenshots together with page and accessibility state to ground actions. Customers do not maintain CSS selectors for the pilot, but the system is not a vision-only human replica. Prompt, model, browser and configuration versions are retained with run provenance.