The code patching preset has six token costs: three hundred fifty, four hundred twenty, two hundred eighty, two hundred twenty, ninety and one hundred sixty. Their sum is one thousand five hundred twenty. Reset initializes total tokens to the first planner cost, three hundred fifty, rather than zero. The first step then adds that planner cost again. With a passing evaluator and no reflection the completed counter is consequently one thousand eight hundred seventy, not one thousand five hundred twenty. Bars compare these values at point one two pixels per token. These are authored simulator costs, not tokenizer measurements, provider pricing or actual model calls. The default mock score point seven eight is below the point eight five threshold. At the evaluator the code raises the score by point one two to point nine and jumps back to the tool index. Replaying tool and evaluator adds two hundred eighty plus two hundred twenty tokens, while reflection adds another one hundred eighty. The extra six hundred eighty takes the preloaded counter from one thousand eight hundred seventy to two thousand five hundred fifty. Bars compare those totals at point zero eight pixels per token. Native offline stepping confirms two thousand five hundred fifty and finished status. The telemetry update precedes the reflection penalty, so the visible counter temporarily lags internal state until the next step. The score improvement is assigned arithmetic, not learned quality or actual test execution. The failure checkbox adds one hundred fifty tokens and four hundred latency units, clears itself and advances to the next node. Despite the retry log wording, that branch does not replay the tool step. A default reflective run with this injection therefore totals two thousand seven hundred tokens versus two thousand five hundred fifty without it, drawn at point zero eight pixels per token. Each normal step independently assigns random latency between two hundred and four hundred ninety nine milliseconds at speed one. This is not observed execution time. Missing Cytoscape uses a healthy native SVG fallback with six node boxes but no drawn edges; JSON exports actual local counters and SVG exports the static preset boxes. No real human approval, pull request, external tool or model request occurs. Paired native controls and downloads preserve these limits rather than turning fixture captions into empirical claims.