Why late requests can increase while average elapsed time falls

Original fictional teaching exercise. Not a user result or an approved plan.
Keep this pack with example.json; all numerical facts are shared between languages.

CSV FIELDS AND SCOPE
period: earlier is September 7–13, 2026; later is September 14–20. Both are fictional observation windows.
request_id: a unique exercise request identifier; periods are not paired tracking of the same requests.
elapsed_minutes: minutes from receipt to resolution, not staff effort. Each period has ten completed records, with no open, excluded or missing cases in this exercise. The preset late threshold is strictly greater than 20 minutes.

BLANK WORKSHEET
Receipt and resolution events, unit:
Period windows, samples and open records:
Duplicates, missingness and pairing:
Source records and file checksum:
Counts, means, medians and distributions:
Threshold, late counts and denominators:
Composition, comparability and causal limits:
Report statement and next check:

PRACTICE
Put elapsed-times.csv and recalculate.py together; run python3 recalculate.py. Verify means 10 and 8.4, medians 8 and 4, late counts 1 and 2 (>20 minutes). Rewrite an unsupported claim of overall improvement.

CHECK
Retain units, periods, denominators and unknowns. Distinguish a proposal, approval, action completion and observed effect.