No magic.
Just inspectable signals.

Two ways to look under the hood

This deployment inspects static HTML only. Browser rendering is unavailable here, so JavaScript-generated content and computed appearance are not inspected. Reports identify this limitation.

Every scan starts with public HTML. When JavaScript or external styles could add useful evidence, an isolated browser also observes the rendered page and its computed styles. Reports label evidence as Static HTML, Rendered DOM, or both.

Rendered findings are scored once, with the initial HTML inspection retained separately for comparison. If rendering fails or capacity is busy, the static report remains available with a clear limitation.

Browser inspection is passive: it does not click buttons, submit forms, or sign in. Explicit no-op controls and selected runtime error patterns can provide context. Missing inline event handlers do not establish that a control is broken. Runtime observations use summarized counts, not raw console messages.

A score you can follow

The Vibe Score is a 0–100 experimental pattern index, not a percentage probability of AI authorship. Several independent patterns matter more than an isolated design choice. Related findings have bounded influence; thoughtful implementation can provide counter-signals.

Category meters describe pattern intensity within each category. They are not shares of the total. Evidence cards explain the observed findings, their source, and their contribution. Model 2.5 is the frozen detection model used by V1.

One gradient, rounded card, em dash, framework, or icon library cannot tell us who built a site. Familiar visual composition becomes more informative when it appears alongside independent implementation, content, or starter-template evidence.

Confidence measures coverage

Confidence describes how much useful material was available to inspect, independently of the Vibe Score. A sparse page, failed resources, or an incomplete browser inspection reduces coverage.

High confidence means substantial content was inspected. It does not mean we know whether AI was involved. A low score with high coverage is a valid result.

The limits matter

Human developers use templates. AI-assisted developers can produce thoughtful, accessible code. Both false positives and false negatives are possible. The supplied real-world examples did not establish reliable separation between development workflows.

Only one public page is observed, at one desktop viewport and during a short observation window. Login walls, late-loading content, hidden elements, shadow DOM, blocked assets, and English-centric copy rules can limit findings. There is no screenshot interpretation, image recognition, LLM, or AI text detector.

Synthetic calibration checks relative behavior across representative styles. It is not a labeled real-world accuracy study. Treat the evidence as a prompt for inspection, never a judgment of a developer.

Public pages, bounded requests

Only public HTTP and HTTPS pages are supported. Requests, redirects, browser resources, execution time, and process capacity are restricted. Private networks and credential-bearing URLs are blocked. Unexpected failures return safe messages.

Your report stays in this browser tab for up to an hour. Public report links are unavailable. Share the analyzer homepage so someone else can run their own scan, or share the clearly labeled example. There is no account, server-side report storage, or public history.

Raw fetched HTML is processed transiently. Analysis results are returned to your browser and are not written to server files. Hosting logs have their own retention policies. Closing the tab, expiry, or unavailable browser storage can remove access to a private report.