C04 Solutions

Reference answers for ../exercises/. Read them after writing your own — the value is in the reasoning and the tradeoff commentary, not the conclusions.

#	Solution	Headline
1	01_optimization_order_solution.md	Two different problems wearing one hat. Caching (once the key is fixed) solves ~31% of the cost; nothing on the cheap rungs solves the latency SLO, and distillation is wrong here because the base model dies in two quarters.
2	02_quality_gate_solution.md	The proposed evidence is four separate failures at once. The safety-critical queues are 0.9% of volume, so the uniform eval set is structurally incapable of detecting a regression there — that is the whole problem to solve.
3	03_build_the_cost_case_solution.md	Cache + batching delivers ~50% for 5 engineer-weeks and survives the platform migration. Distillation loses ~$24,500 against a model with ~7 months to live: decline it.

How these are written

Assumptions stated, arithmetic shown, tradeoff named, "it depends on X" pivot given, next measurement proposed. Numbers are either given in the problem or derived; where one comes from a ../tools/ model, that is labelled — a model output is not a measurement, and the multipliers in those models are placeholders.

Grading yourself

The gap between a competent answer and a strong one, in this module specifically:

• A competent answer picks the right optimization. A strong one computes the ceiling first and eliminates candidates before analysing them.
• A competent answer mentions evaluation. A strong one gives numeric thresholds per slice and a rollback trigger an on-call engineer could execute without judgement.
• A competent answer recommends action. A strong one recommends declining something, with the number that justifies it.
• A competent answer treats cost and latency as one problem. A strong one notices they are two, with different answers, and that some fixes trade one against the other.

If your answer would have approved everything on the list, redo it.