Module 8 practice · keep the measurements separate

Attempt using the lesson, then compare the solutions.

A. Compute from explicit labels. Three claims c1,c2,c3 all cite existing evidence IDs. Manual support labels are true, false and true. The required aspects are c1,c2. Calculate valid-citation-ID count, supported-claim fraction and required supported-claim coverage. Explain what an ID-existence check misses. Use review_claims to verify the arithmetic after your calculation.

B. Refuse a misleading perfect score. An answerable question requires two aspects. The answer is empty. What should proposed-claim support fraction and required coverage be? Why should a system not report 100% support simply because it made no unsupported claims?

C. Repair a split. Ten question rewrites about the same renewal rule are split randomly, five into development and five into test. The author has tuned synonyms after inspecting development results. Explain why the test may overstate generalization and propose a better split/review procedure. State what kind of evidence the current supplied fixtures actually provide.

D. Diagnose one trace. Retrieval recall is1, packet evidence coverage is0, and the template abstains. Which stage should be investigated first? In another case both coverage measures are1 but a human labels an unconditional renewal claim unsupported. What should be inspected next? Explain why changing the retriever in both cases would obscure the diagnosis.

Self-review: correct denominators and null handling, no semantic credit from IDs alone, a meaningful split and a repair matched to the failing stage.