C03 Solutions

Reference answers for ../exercises/. Read them after writing your own — these include the reasoning and the tradeoff commentary, which is the part that transfers.

#	Solution	Headline
1	01_diagnose_slow_training_solution.md	Two overlapping problems: input-pipeline starvation from the re-export, and a straggler tail that grows with rank count. The loss bump at restart is a third, separate bug.
2	02_checkpoint_strategy_solution.md	T* = 65.7 min synchronous → 5.2 min sharded+async, overhead 14.8% → 3.6%. Spot wins on cost by ~54%, and the deciding variable is restart time, not the checkpoint interval.
3	03_repro_audit_solution.md	The +1.3 claim is unsupported, not merely unverified: the eval set changed and there is one seed. Fix the comparison before fixing determinism.

How these are written

Each solution follows the structure the exercises ask for: assumptions stated, arithmetic shown, tradeoff named, "it depends on X" pivot given, and a next measurement proposed. Where a number appears it is either given in the problem or derived — and where it comes from one of the ../tools/ models, that is labelled explicitly, because a model output is not a measurement.

A note on grading yourself

The most common gap between a competent answer and a strong one is not knowledge — it is ordering and honesty:

• A competent answer lists the right causes. A strong one lists them in the order it would actually check them, and says why that order.
• A competent answer recommends improvements. A strong one names the thing it would decline to do and the money that saves.
• A competent answer gives a number. A strong one gives the number, its sensitivity, and the fact that would change it.

If your answer had no numbers, or recommended everything, redo it before reading further.