Read the explanation
The overhead model divides two hundred fifty lines by thirty five times familiarity point eight, yielding about eight point nine two nine manual hours. Assisted time sums four constructed stages: prompting point five eight, inspection two point three four three seven five, context repair one point five and debugging three point two six two five, total seven point six eight six two five hours. Bars encode both at thirty pixels per hour. These equations use authored productivity assumptions rather than measured developer time or model performance. The source comment mentions a fifty-line baseline but its actual formula uses thirty five. The speed ratio is manual divided by assisted, about one point one six, and does not establish an empirical tool advantage. Changing only prompt iterations from four to twelve increases prompting by point nine six hours and context repair by three hours. The inspection and debugging terms stay fixed, so total assisted time becomes eleven point six four six two five, while manual time remains eight point nine two nine. Bars compare default and changed assisted time at thirty pixels per hour. Ratio becomes about point seven seven, marking overhead drag in the source. That ratio below one is throughput relative to the manual baseline; interpreting point seven seven times slowdown literally would invert the duration relationship. Native inline slider handlers compute captions before D3 throws, so metrics and actual JSON export work while diagrams fail offline. No real model runs or coding tasks are timed. Verification tax is defined broadly as inspection plus context repair plus debugging divided by all assisted time. Defaults yield about ninety two point four five percent; twelve prompt iterations give about eighty six point seven eight even though absolute verification hours increase from seven point one zero six to ten point one zero six. Bars encode percentages at three pixels per point. A lower fraction therefore need not mean less work. The diagram applies minimum forty-pixel stage widths that can exceed true proportional encoding for small stages, and the displayed charts use synthetic rather than observed times. Native changed control, actual state-and-analysis JSON, four viewport transitions, positive original canvas buffers and all footer placement are preserved locally, without public publication or exhaustive workflow or cost validation.