Read the explanation
The written expectation sums the probability of reaching each allowed iteration. With convergence chance point eight five and maximum three, it is one plus point one five plus point zero two two five, or one point one seven two five. Raising only chance to point nine five gives one point zero five two five. At two hundred pixels per expected iteration the bars span two hundred thirty four point five and two hundred ten point five. Despite the input label max retries, the simulation bound includes the first iteration, allowing at most three total attempts. These probabilities are user assumptions and random flags, not measured reliability of actual agent-generated outputs. The self-correction fixture node latencies sum to twenty five hundred milliseconds and average token fields to sixteen hundred fifty. Multiplying by expected iterations one point one seven two five gives expected fixture latency twenty nine hundred thirty one point two five, represented by two hundred and two hundred thirty four point five pixel bars at point zero eight pixels per millisecond. The token cost constant point zero zero zero zero zero two gives expected cost about point zero zero three eight six nine two five. These are arithmetic over authored inputs, not actual provider charges, measured token usage, or wall-clock execution. No model or external API is called by this local loop. The field labeled p ninety five multiplies per-iteration latency by the smaller of expected iterations times one point five and the allowed count. For this fixture it rounds to forty three hundred ninety seven milliseconds. At point zero six pixels per millisecond the bars compare expected twenty nine hundred thirty one point two five with that caption, spanning one hundred seventy five point eight seven five and two hundred sixty three point eight two. This is a heuristic multiplier, not a sampled ninety-fifth percentile or a derived discrete quantile. The offline Cytoscape dependency is absent, so loading presets returns and metric recomputation throws when it accesses null graph nodes. Native slider labels can change, but a healthy renderer, numeric outputs, graph export and actual loop execution remain unproved.