Modules C04–C05, C07 · concept player

Cheapest and least risky fix first

Four levers, applied in the module's stated order: stop calling the model at all, serve it better, shrink the numbers, shrink the model. Each is more invasive and more risky than the last — slide them and watch how rarely you need to reach the expensive end.

1· Cache / cascade hit-rate (lowest risk): 30%
2· Serve better — batching / runtime / graph capture: 2.0×
3· Shrink the numbers — quantization: 2.0×
4· Shrink the model — distillation (highest risk): 1.0×
Target cost budget: 25% of baseline
100%cost after all four levers
nonefirst stage that clears budget
Lowrisk of the levers you actually needed

Remaining cost, stage by stage

These are levers you set, not measured savings. The multiplication is exact given the factors above; what any real workload's cache hit-rate or quantization speedup actually is has to come from your own traffic and your own profiling, not this page. The ordering is the citable part — C04 section 1: cheapest and least risky first, because a lever you do not need is a lever whose risk you never took on.