This explains the saved replay demonstration, not observed security test results. The evaluator scans stored steps from the first through the currently selected step. It stops at the first step whose guardrail check names an enabled checkbox. If a hypothetical enabled guard matches step two and another matches step four, the reported block is step two, not the latest or strongest mechanism. With eighty pixels per step position, bars compare one hundred sixty and three hundred twenty pixels. The first match follows array order. If no earlier match exists, a checked prompt sanitizer adds a fallback block at step two once the replay has reached that position. This demonstration describes chosen source rules; a checkbox does not enforce a real sandbox or prove that an external action was prevented. The displayed risk switches between fixed strings: eighty eight percent when unblocked and fourteen percent when blocked. The page also writes a fixed eighty four percent reduction label. Subtracting fourteen from eighty eight gives seventy four percentage points; dividing that change by eighty eight gives about eighty four point zero nine percent relative reduction. Bars compare the stored risk numbers at three pixels per percentage point, measuring two hundred sixty four and forty two pixels. That numerical relationship does not establish measured risk or compromised asset reduction. The source does not count actual system assets for that label. Its blast radius when blocked is block step minus one followed by a system label, while an unblocked replay uses the stored step blast text. Those are authored telemetry conventions, not an independent security measurement. A blocked step of two produces a displayed count of one system; a blocked step of four produces three. At one hundred pixels per displayed system count, those bars measure one hundred and three hundred pixels. The arithmetic comes from step position, not traversal of a discovered infrastructure graph. The export writes selected scenario, its attack vector text, a literal five replay steps, the computed block position, risk strings, mechanism and guardrail flags. It does not export actual probe responses or the complete rendered log. Moving the timeline or changing guards recomputes a scripted state. The local original depends on D three for its graph, and the unavailable offline library limits verification of replay visuals. Native control and export checks must preserve the original behavior, including any source errors, without treating matching failures as functional completion.