Actual screen bytes
Two 1348 × 926 JPEGs and SHA-256 digests preserve the top and scrolled states. The DOM observation records visible links, labels, destinations, and locations.
A real Portland security page repeats its call path in a banner, navigation, and fixed dock. An ĀML candidate hides one repetition. A deliberate over-suppression looks fine at the top but fails after the banner scrolls away.
Production screen, local counterfactual. The screenshots are direct captures of the live page. Colored outlines and candidate decisions are an explanatory overlay. The production website was not modified in this experiment.
Reading checked-in report.
The counts cover these two captured desktop viewports. They do not assert behavior at other sizes, times, routes, or assistive-technology states.
Two 1348 × 926 JPEGs and SHA-256 digests preserve the top and scrolled states. The DOM observation records visible links, labels, destinations, and locations.
Attention cost and restoration value are project-authored design inputs, not measurements of visitors. The candidate keeps the fixed call and email dock while suppressing the repeated navigation number.
The adversarial probe also suppresses the dock call. The top still has the banner; the scrolled state loses every observed call path, so the continuity gate fails.
From the repository root, rebuild the visual report and verify both source screenshot digests. The test checks the passing candidate, failing scrolled counterexample, and malformed labels.
node scripts/build-field-contact-continuity-report.mjs --check
node --test test/field-contact-continuity.test.jsThe live page may have changed since capture. A matching hash verifies the checked-in image bytes, not the truth of the labels or an improvement in customer outcomes.