The explorer normalizes each driver by subtracting its minimum and dividing by its range, clipping the result between zero and one. Research Agent uses eight steps on a one-to-twenty range, giving seven nineteenth, about point three six eight. Two tools on a zero-to-eight range give point two five. Both bars use six hundred display pixels per normalized unit. Steps are labeled low, while tools are labeled none, because category comparisons include their upper boundaries: through point two five is none, through point five low, through point seven five medium, and above that high. None therefore does not necessarily mean a zero input. These labels describe positions within configured ranges, not measured costs. Each driver node's radius is sixteen plus twenty times its normalized value. Zero gives radius sixteen and one gives thirty-six, shown at fifteen display pixels per source radius unit. Its initial distance from the base is one hundred plus eighty times the same normalized value, ranging from one hundred to one hundred eighty. The visual area grows quadratically with radius, so twice a radius does not mean twice a cost. D3 forces can move nodes after their initial placement. Neither the radius nor distance comes from dollars, measured inference time or token billing. The graph provides qualitative visual emphasis for an assigned driver score. The graph always starts with an inference base and four driver nodes. Each chosen tool adds one small node linked to reasoning. Research Agent's two tools therefore produce seven nodes; Code Agent's four tools produce nine. Both counts share fifty display pixels per node. Tool nodes start sixty plus twice the steps away from the reasoning node, so eight steps assign distance seventy-six and twelve assign eighty-four. Presets also change context and memory inputs. Export records state and normalized driver labels, not an invoice or observed agent trace. The saved external D3 tag has an authored integrity string, and offline checks block external dependencies; native behavior must be compared separately. This narration explains saved source arithmetic without executing a real model or charging any backend.