Download the complete part-02 practice pack (ZIP). Extract it and preserve its folders. All organizations, records, and results in this teaching pack are fictional.
Use these forms with Chapters 6–9. Adapt the fields to the actual task. The supplied Mesa documents, job records, questions, and scores are fictional teaching material. Blank observed-result fields are intentionally unfilled.
Data preparation brief
Business question:
Unit being counted or measured:
Included records:
Excluded records:
Date field and reporting period:
Source system, extract time, and scope:
Required fields and definitions:
Known missing or stale information:
Possible duplicates:
Ambiguous formats or labels:
Source owner and correction authority:
Unchanged original location:
Working copy location:
Correction log:
Approved calculation/filter rules:
Unresolved limitations attached to the result:
For the Chapter 6 exercise, inspect jobs-raw.csv before opening the resolution notes. Write the questions you would ask the owner. Then use Owner-Resolutions.md to produce your own resolved view and compare it with jobs-resolved.csv.
The expected count is four distinct completed inspections. The known eligible fee subtotal is $135, with one missing fee. The exercise does not establish collected revenue. The verification script checks these fixture calculations; it does not perform an AI evaluation.
Correction log
| Source row IDs | Observed issue | Proposed change | Owner evidence/decision | Status | Effect on result |
|---|---|---|---|---|---|
| Complete for your records |
Keep “possible duplicate” and “confirmed duplicate” distinct. Do not resolve a conflict by silently choosing the value that makes the report look complete.
Prompt brief and change record
Prompt ID and version:
Task and work product:
Audience and purpose:
Approved sources:
Input fields:
Allowed output statuses and their meanings:
Unknown/missing-information behavior:
Action boundaries:
Review rubric:
Development case IDs:
Held-back evaluation case IDs:
Tool/model/configuration:
Change from previous version:
Reason for change:
Observed comparison results:
Unresolved failures:
Decision and owner:
“Ready for human review” is a description of a draft, not permission to execute a business action. Define the later approval separately.
Evidence review
| Answer claim | Cited source and location | Source exists? | Passage supports claim? | Applies to this case/date? | Condition omitted? | Review action |
|---|---|---|---|---|---|---|
| Complete for each material claim |
Use the current documents in practice/sources/ for ordinary current-policy questions. The archived document is supplied separately for a deliberate historical/conflict test. It must not be treated as current merely because it is present in the package.
For a missing warranty term, the expected response is to identify the gap. A citation to the inspection policy does not supply the missing warranty.
Retrieval investigation record
Question and user role:
Authorized source set:
Source versions and index refresh time:
Evidence actually supplied to the model:
Was the needed evidence present?
Were relevant conditions preserved?
Did unauthorized or superseded material appear?
Answer and cited passages:
Observed failure:
Likely stage to investigate:
Correction and owner:
Cases to rerun:
Text-only practice cannot establish that a real application's permission checks work. Keep application access tests distinct from prompt-answer tests.
Evaluation plan
Objective and intended scope:
Rubric and overall pass rule:
Critical failure definitions:
Case sources and selection method:
Case categories and counts:
Development/evaluation separation:
Configurations being compared:
Number of runs and reason:
Scoring method and reviewer:
How disagreements are resolved:
Automatic checks and what they cannot establish:
AI-grader validation, if used:
Quality measures and denominators:
Time/cost measures and included costs:
Release or continuation conditions:
Observed results:
Known gaps:
Decision:
Next review triggers:
The score tables in Chapter 9 are invented examples to practice interpreting results. They are stored with explicit illustrative labels. Use the blank evaluation log from Part 1 for actual observations.
Run the example checks
If Python 3 is available, open a terminal in the practice directory and run:
python3 verify_examples.py
The script reads only the bundled fixture files. It checks record counts, approved normalization, arithmetic, source-file references, and the illustrative scoring totals. It makes no network request, calls no AI model, and writes no business records. Passing these checks means the published practice material is internally consistent; it does not validate the behavior of an AI application.