MODEL-TOOL EXECUTION GRAPH (DAG)
Normal Step
Latency Bottleneck
High Reliability
CONTEXT TOKEN SATURATION & SLIDING DECAY
The default workflow uses vector search, code interpreter, SQL and code interpreter again. Their assigned latencies one hundred twenty, eight hundred fifty, three hundred forty and eight hundred fifty sum two thousand one hundred sixty milliseconds. A special fixture override instead reports two thousand four hundred ten, difference two hundred fifty. At point one pixels per millisecond computed steps measure two hundred sixteen, fixture two hundred forty one and gap twenty five. Web search is defined in the tool library but unused in these steps, even though the default recommendation names it. This local arithmetic and override are not actual model or tool benchmark measurements. Four step costs sum point zero zero two plus point zero three plus point zero zero five plus point zero three, or point zero six seven dollars. The fixture reports point zero four seven, two cents less. At four thousand pixels per dollar the computed sum bar is two hundred sixty eight, fixture one hundred eighty eight and gap eighty. Likewise multiplying assigned reliability point ninety nine, point ninety eight, point ninety five and point ninety eight gives point nine zero three two five six two, while the override reports point eight four two. The code uses fixed values and has no model or API calls. Neither price nor reliability is current provider evidence. Without context decay, step input plus output values are sixteen hundred, thirty two hundred, twenty one hundred fifty and forty three hundred. Cumulative total reaches eleven thousand two hundred fifty, but the fixture peak says five thousand one hundred, a gap six thousand one hundred fifty. At point zero two pixels per token the computed peak is two hundred twenty five, fixture one hundred two and gap one hundred twenty three. Turning decay to ten percent computes a different peak nine thousand nine hundred ninety three by repeatedly retaining ninety percent of previous context. Native presets, sliders, reset and actual JSON export update assumptions before missing graph dependencies throw. Run Benchmark only recomputes these formulas; no tool execution, measured latency, actual context memory or production savings is proved.