A wrong answer can begin long before the model speaks. In this guided practice, follow the information path and decide what evidence would change the result. Try each question before its explanation. Use the worksheet to keep your reasoning; the separate answer key remains available afterward. All numerical service cases here are hypothetical. An order total of $100 appears
after joining the order to 4 order lines. A naive sum reports $400. What failed? The join repeated an order-level total across 4 line rows. The order still costs $100. Count or aggregate at the intended grain before combining values. 3 measurements are 4, 8, and missing. What is the mean of known values?
What assumption yields 4? The known values total 12, and there are 2 of them, giving 6. Replacing the missing value with 0 changes the divisor to 3 and produces 4. That 2nd result depends on a replacement assumption. A record is valid JSON with a plausible date but belongs to the wrong customer.
Which quality dimension failed? The record passed some structure and format rules. Its customer relationship disagrees with the relevant reference, so accuracy failed. The wrong relationship may also expose information to the wrong person; investigate access as well as the value. A policy PDF is indexed for retrieval.
Did that alone train model weights? Indexing supplies another route to evidence at request time. Changing learned parameters requires a separate learning process. Test which authorized passage is retrieved and whether the answer uses it correctly. A model generates 1000 labeled examples and judges them correct. Is this independent ground truth? The generator judging its own output
does not provide independent ground truth. Recompute arithmetic, compare policy answers with the applicable source, or use qualified independent review and real target cases, as the task requires. A source row is deleted but a cached answer remains. Has every representation been removed? The cache remains an affected copy. Follow the data map through source records, indexes, chunks, caches,
logs and other representations within the request's scope. Verify each action and report unresolved retention or access paths. Now connect the stages. In the course's hypothetical 10-question pilot, 8 questions retrieve the required authorized policy evidence. 7 of those 8 receive correct explanations. End-to-end evidence-supported correctness is 7 divided by 10, or 70%.
Generation conditional on successful retrieval is 7 divided by 8, or 87.5%. Reporting only that 2nd number hides the 2 retrieval failures. Count a disclosure of unauthorized information separately as a high-consequence failure, even when other answers are correct. A hypothetical 12-question test retrieves required authorized evidence for 9 questions.
6 of those 9 receive correct evidence-supported explanations. What are the end-to-end and conditional-generation rates? How should an unauthorized disclosure be reported? 6 correct evidence-supported answers across 12 questions give 50% end-to-end. Among the 9 successful retrievals, 6 correct explanations give about 66.7%. Retrieval coverage is 9 of 12, or 75%.
These figures describe different stages and denominators. Any unauthorized disclosure still needs its own incident and access response. For the independent task, choose 1 upstream assumption from your own example. Name its consequence, the evidence that could resolve it, and what remains unknown. Across the questions, the same discipline made the difference:
preserve what 1 record means, keep evidence roles separate, and inspect the complete path before trusting a result. Carry that reasoning into your data-architecture project.
Main Point
Explain what the inputs and processing steps support before accepting an answer.
Key Points
Preserve row grain and distinguish unknown values from measured zero.
Separate format checks, factual accuracy, permission and evidence roles.
Measure the whole path with its full denominator.
Try the seven questions before opening their explanations. Use the worksheet to keep your reasoning and the separate key to check it.
1. Worked stage measurement
Now connect the stages. In the course's hypothetical ten-question pilot, eight questions retrieve the required authorized policy evidence. Seven of those eight receive correct explanations. End-to-end evidence-supported correctness is seven divided by ten, or seventy percent. Generation conditional on successful retrieval is seven divided by eight, or eighty-seven point five percent. Reporting only that second number hides the two retrieval failures. Count a disclosure of unauthorized information separately as a high-consequence failure, even when other answers are correct.
2. Independent transfer
Use Q07 to inspect a distinct twelve-question case. Then complete P01 with a specific upstream assumption, corrective evidence/control and residual uncertainty.
3. Bringing the practice together
A dependable explanation identifies the record unit, the information actually available and the process that changed it. Your architecture can now show where a wrong answer begins and what evidence is required before the affected use resumes.
Learning Checks
Attempt each answer before expanding the reasoning.
M13-PRACTICE-Q01
An order total of 100 dollars appears after joining the order to four order lines. A naive sum reports 400 dollars. What failed?
Show reasoning
The join multiplied an order-level value across finer-grain rows; aggregate or join at the intended grain.
M13-PRACTICE-Q02
Three measurements are 4, 8, and missing. What is the mean of known values? What assumption yields 4?
Show reasoning
Known-value mean is 6; treating missing as zero yields 4 and imposes an assumption.
M13-PRACTICE-Q03
A record is valid JSON with a plausible date but belongs to the wrong customer. Which quality dimension failed?
Show reasoning
Accuracy and possibly permission context failed; format validity alone did not establish correct assignment.
M13-PRACTICE-Q04
A policy PDF is indexed for retrieval. Did that alone train model weights?
Show reasoning
No; retrieval availability changed, not model parameters.
M13-PRACTICE-Q05
A model generates 1,000 labeled examples and judges them correct. Is this independent ground truth?
Show reasoning
No; use external constraints, qualified review or independent target evaluation.
M13-PRACTICE-Q06
A source row is deleted but a cached answer remains. Has every representation been removed?
Show reasoning
No; trace copies/indexes/caches and other in-scope derived representations and verify actions.
M13-PRACTICE-Q07
A hypothetical twelve-question test retrieves required authorized evidence for nine questions. Six of those nine receive correct evidence-supported explanations. What are the end-to-end and conditional-generation rates? How should an unauthorized disclosure be reported?
Show reasoning
End-to-end 6/12=50%; conditional generation 6/9≈66.7%; retrieval coverage 9/12=75%. Report any unauthorized disclosure separately as a high-consequence failure; never average it away.
For the synthetic-label check, the public guidance below explains why generated examples require independent reference and target-evaluation checks. It does not validate the hypothetical 1,000 cases. [UK-SYN]
Sources
PG-TABLE: PostgreSQL Global Development Group, PostgreSQL 18 Documentation, §7.2, “Table Expressions,” qualified joins and WHERE. https://www.postgresql.org/docs/18/queries-table-expressions.html. Checked 29 October 2026.
PG-AGG: PostgreSQL Global Development Group, PostgreSQL 18 Documentation, §9.21, “Aggregate Functions,” AVG/COUNT entries. https://www.postgresql.org/docs/18/functions-aggregate.html. Checked 29 October 2026.
UK-FRAME: UK Government, *The Government Data Quality Framework*, 3 December 2020, principles and dimensions. https://www.gov.uk/government/publications/the-government-data-quality-framework/the-government-data-quality-framework. Retrieved 29 October 2026. Public-sector guidance, not a universal certification standard.
AZ-RAG Microsoft. *Retrieval-augmented generation in Azure AI Search*. https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview. Live architecture documentation; revision date not asserted. Retrieved 29 October 2026.
ICO-ERASURE Information Commissioner’s Office. *Right to erasure*. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/individual-rights/right-to-erasure/. UK guidance under review after legal changes; publication date not asserted. Retrieved 29 October 2026.
UK-SYN UK Government. *AI Insights: Synthetic Data*. https://www.gov.uk/government/publications/ai-insights/ai-insights-synthetic-data-html. Updated 3 August 2026. Retrieved 29 October 2026.
The business scenarios and numerical counts are hypothetical teaching examples. These public references support the underlying methods; they do not report results for the fictional service.
Use the supplied fictional cases and templates for practice.
Practice
Write your answer, compare it with the model answer and decide what to revisit. These written checkpoints are self-assessed.
An order total of 100 dollars appears after joining the order to four order lines. A naive sum reports 400 dollars. What failed?
Show The Model Answer
The join multiplied an order-level value across finer-grain rows; aggregate or join at the intended grain.
Compare your reasoning with the model answer. These written checks are self-assessed.
Three measurements are 4, 8, and missing. What is the mean of known values? What assumption yields 4?
Show The Model Answer
Known-value mean is 6; treating missing as zero yields 4 and imposes an assumption.
Compare your reasoning with the model answer. These written checks are self-assessed.
A record is valid JSON with a plausible date but belongs to the wrong customer. Which quality dimension failed?
Show The Model Answer
Accuracy and possibly permission context failed; format validity alone did not establish correct assignment.
Compare your reasoning with the model answer. These written checks are self-assessed.
A policy PDF is indexed for retrieval. Did that alone train model weights?
Show The Model Answer
No; retrieval availability changed, not model parameters.
Compare your reasoning with the model answer. These written checks are self-assessed.
A model generates 1,000 labeled examples and judges them correct. Is this independent ground truth?
Show The Model Answer
No; use external constraints, qualified review or independent target evaluation.
Compare your reasoning with the model answer. These written checks are self-assessed.
A source row is deleted but a cached answer remains. Has every representation been removed?
Show The Model Answer
No; trace copies/indexes/caches and other in-scope derived representations and verify actions.
Compare your reasoning with the model answer. These written checks are self-assessed.
A hypothetical twelve-question test retrieves required authorized evidence for nine questions. Six of those nine receive correct evidence-supported explanations. What are the end-to-end and conditional-generation rates? How should an unauthorized disclosure be reported?
Show The Model Answer
End-to-end 6/12=50%; conditional generation 6/9≈66.7%; retrieval coverage 9/12=75%. Report any unauthorized disclosure separately as a high-consequence failure; never average it away.
Compare your reasoning with the model answer. These written checks are self-assessed.