# Chapter 12 · Draft answer checks

These answers use constructed data and remain under technical and author review.

1. The pooled rate is $21/100=21\%$. The six batch rates are $10,20,0,20,10,30\%$, whose unweighted mean is $15\%$. The first weights orders; the second weights batches.
2. A copied B06 violates batch-key uniqueness. If wrongly accepted, pooled rate $33/140\approx23.57\%$ and batch mean $1.2/7\approx17.14\%$.
3. `validate="one_to_many"` should reject repeated keys on the left. Without validation, A's three messages multiply both A rows and its outcome weight.
4. Six matched message rows and one unmatched D row produce seven left-join rows. Aggregating messages to case grain and left joining gives four case rows.
5. $42>40$ violates the subset count contract $0\leq L\leq N$. Zero would claim no late orders and conceal an input fault.
6. Excluding an immature pending case yields a mature-cohort rate. Treating it as failure claims an observed failure that did not occur; report pending separately.
7. Total future messages leak events after opening. At a 24-hour review, messages recorded by that review time could be used if the intended deployment has the same information.
8. Whole-dataset fitting leaks held-out information. Split first, fit the scaler on training data and apply those parameters to both partitions.
9. A row split partly tests recognition of previously seen devices. Hold out entire devices to test a new-device deployment.
10. Restart the kernel and run all cells in displayed order, then save source identity, environment and outputs.
11. A manifest can state question, source and hash, row/field count, grain, ranges, code revision, package versions, command, formulas, outputs and limitations. A hash checks bytes; coverage and measurement require evidence.
12. Both programs can share an incomplete export or wrong definition of late. Agreement is useful for detecting language-specific implementation drift.
