The saved code demo selects among four prewritten role presets and executes the chosen JavaScript locally into a preview container. The initial marketing preset displays one thousand two hundred fifty leads, four hundred twenty qualified leads and eighty five deals. At one pixel per five displayed items those bars measure two hundred fifty, eighty four and seventeen. The conversion ratio four hundred twenty divided by one thousand two hundred fifty is thirty three point six percent, and eighty five divided by four hundred twenty about twenty point two four percent. Those arithmetic ratios explain fixed labels; they are not measured campaign results. No model creates new code from the entered specification in this implementation. The trace always appends four records: prompt analysis, code emission, a success or failure label, and final render. The iteration slider changes a counter from one to five, but does not change that four-record trace length. Assigned token use is the floor of eight hundred plus a random value below six hundred, multiplied by selected iterations. At three iterations the range is two thousand four hundred to four thousand one hundred ninety seven. At one pixel per twenty assigned tokens the bars measure one hundred twenty and two hundred nine point eight five. These are random display counters rather than provider usage, and context-budget selection is stored without constraining code execution. Inject Error sets a boolean, attempts to replace the first return keyword, and displays a fixed syntax-error panel. A preset without that keyword may keep its code unchanged while the flag still forces the error view. Run clears the flag and replays the current preset rather than repairing code through a model. Syntax compliance is one hundred or zero based on that same flag, not a separate static-analysis test. At one pixel per percent those two bars measure one hundred and zero. JSON export contains stored specification, current code and assigned telemetry, while edited textarea content is not read back into the stored specification by the run handler. Native preview and download checks can prove local behavior without proving authored benchmark claims.