The saved controls default to twenty-five percent jitter, fifteen percent drift and three retries. Running with fallback produces three schema errors, one timeout and one guardrail block. Their sum is five, giving ninety-five percent success. Retention is point eight five times success, plus eight point five for fallback, minus point zero five times jitter: eighty-eight. Turning fallback off doubles schema errors and adds two timeouts. The raw failure sum becomes ten, but retries above two cap counted failures at six, giving ninety-four percent success and retention seventy-two point eight. Both bars use five pixels per retention point. The log says one hundred randomized loops, yet the code uses rounded arithmetic, no random samples and no loop execution. These assigned scores are illustrative, not empirical reliability or user retention. With fallback off at default jitter and drift, the failure categories sum to ten. At two retries, the cap is fourteen and success is ninety percent. At three retries, the cap drops to six and success becomes ninety-four percent. Both bars use five pixels per percentage point. This discontinuity is a rule in the formula, not evidence of four recovered attempts. The exported failure categories remain the uncapped counts, so they need not sum to the displayed failure percentage. Latency is eighteen twenty plus twelve times jitter minus fifteen times retries, giving twenty seventy-five milliseconds at defaults. Cost is point zero zero nine eight plus point zero zero zero one times drift, minus point zero zero zero four with fallback: point zero one zero nine dollars. Autonomy is a fixed eighty-one or sixty-four percent. Changing blueprint alters graph structure, but these metrics use the same controls rather than executing its nodes. Reset restores default controls but assigns a separate fixture: ninety-four percent success, eighty-eight point five retention, eighteen twenty milliseconds and point zero zero nine eight dollars. A fresh default run calculates ninety-five percent success, eighty-eight retention, twenty seventy-five milliseconds and point zero one zero nine dollars. The two retention bars share five pixels per point. Fault injection subtracts four success points and three point two retention points with floors; it does not alter failure counts or run the selected graph. Four timed trace callbacks add canned steps, including a fixed refund threshold and fixed state-hash label. Repeated runs or reset do not cancel earlier callbacks, so traces can overlap. JSON exports current controls, metrics and graph; generated TypeScript has placeholder node bodies and unconditional edges, not implemented tools or executable conditional policies. No model, customer refund or backend action occurs in this local explanation. The saved page currently has a JavaScript syntax error before initialization. These formulas are source-code derivations, not successfully executed browser results; the offline preview preserves that failure.