The homepage currently puts the analyzer's submit button below the first viewport on laptop and phones, and describes an unincorporated project as a company. This release puts the working public-PR analyzer first, adds a clearly labeled authored report preview, and identifies Ozan Kenan Güngör as the individual operator across public pages, notices, README and structured data.

The initial view explains public GitHub PR input, Claude review and human verification. The real Rails example only prefills the form. The existing pipeline illustration remains available below the tool, with explicit animation activation and preserved motion/WebGL fallbacks. The three exact publication-approved quotes move to Evidence; historical synthetic failures and founder-run Rails limitations remain disclosed.

Analyzer failures now distinguish invalid URLs from service errors and handle non-JSON platform responses without parsing exceptions. The endpoint rejects cross-site browser submissions and limits each warm instance to two concurrent analyses and six starts per rolling minute. This is a local safeguard, not distributed protection or a global spending cap.

Validation: frozen install, lint, typecheck, 369 unit tests, build, 32 offline evaluation fixtures and git diff --check passed locally. Nineteen Chromium browser checks passed across all six requested sizes, including keyboard operation, scanned accessibility, public routes/metadata/links, example prefill, stubbed reports/errors, timeout, reduced motion, WebGL failure and content without JavaScript. All browser analysis requests are intercepted; no Claude requests or paid evaluations were made. CI now repeats the browser checks on Node 24.

Measured local optimized-build initial desktop JS decreased from 421,725 to 180,657 encoded bytes. The 360×800 submit bottom moved from y=889 to y=717; observed CLS remained zero. Single-run unthrottled lab observations are not Lighthouse scores, field performance or model accuracy. Runtime dependencies and the Claude prompt/model/schema/coverage policy, GitHub retrieval, evaluator and published evidence records are unchanged.

See the audit, measured comparisons, adversarial review and remaining limitations. Platform-wide rate protection, provider spending limits, physical-device performance and live model quality are not established by these offline checks. Production delivery will use the existing Vercel integration after both CI and deployment checks pass.