Day 1 · Monday · Answer key

Prediction claims and credible evaluation

Pre-class concept check

1. P1

   B. The booking and scheduled date exist before discharge; completed follow-up and readmission are later outcomes.

2. P2

   C. Always predicting “on time” scores 92%, so 93% is a one-point accuracy improvement; queue performance is still unknown.

3. P3

   B. Both later time and unseen organisations belong in the test because both belong in the proposed use.

4. P4

   B. A calling decision requires the change caused by contact, which compares contact with no contact for comparable customers.

5. P5

   C. Selective accuracy describes answered cases; coverage states how often the system answers.

6. P6

   B. Once audit errors guide revision, the audit has become development evidence.

Concept studio

1. C1

   A. The unit, moment, outcome, and review action all refer to the same parcel decision.

2. C2

   D. Realised hub delay arrives after the acceptance-time score.

3. C3

   C. Accuracy alone omits prevalence, baseline, capacity, and the location of errors.

4. C4

   B. An untouched latest month most closely represents next-month use among current merchants.

5. C5

   A. Accuracy is \((6+84)/100=90\%\), precision is \(6/10=60\%\), recall is \(6/12=50\%\), and the baseline is \(88/100=88\%\).

6. C6

   A. Six late parcels in the queue is well above the random expectation of 1.2, but the audit does not price actions or guarantee stability.

7. C7

   C. Predictive association with courier does not identify the effect of changing courier.

8. C8

   A. The audit supports assisted prioritisation, not automatic rerouting or a causal story.

Empirical studio

1. E1

   B. The target is supported role extraction from the supplied description, not actual authority.

2. E2

   C. Only the exact phrase in the supplied description belongs to the declared information set.

3. E3

   B. Freezing prevents audit outcomes from selecting the rule they are meant to evaluate.

4. E4

   C. The compact rule answers 245 rows and matches the provisional reference on those 245; the other five are abstentions.

5. E5

   B. Evidence preservation, review, correction, and a path for abstention make the workflow contestable.

6. E6

   A. Two phrase-based systems can agree because they share construction choices; blinded review supplies more independent evidence.

7. E7

   B. The reported result is limited to this frozen build and provisional audit reference until independent review is complete.

Problem set

1. S1

   A. Precision uses flagged parcels as its denominator: \(30/40=75\%\).

2. S2

   C. Recall uses all late parcels as its denominator: \(30/50=60\%\).

3. S3

   A. Thirty late parcels rather than four expected at random is strong prioritisation evidence, not evidence of prevented delays.

4. S4

   C. A week-three appointment cannot be known before term.

5. S5

   B. Outcome prediction does not identify the effect of offering support.

6. S6

   B. The stated use is voluntary, reviewable outreach with attention to errors and burdens.

7. S7

   C. Later documents from fully unseen organisations best match both dimensions of the use.

8. S8

   A. Conditions can still drift between the evaluation period and 2027.

9. S9

   A. Revision on final cases consumes their independence; a fresh final check is required.

10. S10

    B. Coverage is \(95/100\); selective accuracy is \(93/95\).

11. S11

    C. The task requires abstention when the description supplies no evidence.

12. S12

    B. Extracted wording frequency is observed; institutional promotion practice is not.

← Day 1 lecture  ·  All four days →