The saved source describes a planner, two coding agents and an evaluator. Their assigned latencies are twelve hundred, twenty-one hundred, fifteen hundred and nine hundred milliseconds. If everything ran sequentially, the sum would be fifty-seven hundred milliseconds. Under an ideal parallel fork and join, both coders start after the planner and the evaluator waits for the slower coder. The longest route is therefore twelve hundred plus twenty-one hundred plus nine hundred, or forty-two hundred milliseconds. Both bars use one hundred pixels per second. This is a mathematical explanation of the fixture, assuming independent branches and zero scheduling overhead. It is not an observed execution trace or a measured speedup. Now imagine reducing coder A from twenty-one hundred to one thousand milliseconds while coder B stays at fifteen hundred. A genuine longest-path calculation would switch the slower branch to coder B and yield twelve hundred plus fifteen hundred plus nine hundred, or thirty-six hundred milliseconds. But the saved evaluator-optimizer branch does not compute that change: it assigns the original forty-two hundred constant, a fixed critical path and a fixed coder A bottleneck. It also assigns an eighteen point two two five dollar cost fixture. Those labels cannot demonstrate responsive optimization. Other presets sum all node latencies instead of traversing the graph. Treat these as source-defined estimates, not provider pricing or benchmark results. Before interpreting any displayed metric, check whether the page can initialize. Its application script declares a constant named MODEL hyphen PRICING. A hyphen is not valid inside a JavaScript variable name, so the script fails to parse before initialization or handlers become available. The saved footer still says Verified Local Topology Engine, but that static label is not execution evidence. The four assigned token counts sum to four thousand fifty; the bar illustrates those source numbers rather than a real model call. The zero-length lower bar represents no verified run result, not zero cost or zero latency. This local explanation preserves the original source, identifies the failure honestly and does not execute generated LangGraph code, contact models or repair production.