Governed analytics sandbox (clean)
A read-only query log against the governed analytics sandbox: run_id, source_system, batch_id, event_time, ingested_at, entity_id, metric, value, schema_version. Used for the Guided Lab.
DS 302 Practice Center
Use the synthetic governed analytics sandbox to practice deciding when an agent is justified, critiquing an analysis plan, designing tool contracts, resolving a semantic metric, testing a permission boundary, evaluating golden questions, defending against prompt injection, and drilling a verification harness.
Info
DS 302 practice data is synthetic teaching data. Every source system, batch, and value is illustrative and does not represent any real organization or individual.
Warning
Only use the synthetic files provided by the course for practice. Unless authorized, never paste confidential organizational data, personal data, or sensitive information into an AI tool or practice environment; check your organization’s policy before using real operational data.
Download the practice files
Governed analytics sandbox (practice, with challenges)
Same field structure, with deliberately designed challenges: late-arriving events, one schema-drift value, a duplicated run, and at least one row that violates the data contract. Used for the Practice Center and the Independent Lab.
Data dictionary (Traditional Chinese)
Field types, units, grain, allowed values, and provenance for the analytics sandbox (Traditional Chinese).
Data dictionary (English)
English data dictionary; fields and definitions match the Chinese version.
Guided Lab starter notebook (verifiable read-only agent)
A starter notebook with guidance and tasks to complete, practicing plan → read-only query → dataset-version citation → verification.
Independent Lab starter notebook / template
Starter notebook and checklist template for the Data Agent Architecture Canvas + Governed Prototype + Evaluation Pack.
Practice tasks
A: Agent-vs-Workflow Decision
Given a set of sample requests, decide which genuinely need a dynamically planning agent and which are better served by a fixed dashboard or workflow.
Files used: Governed analytics sandbox (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Read three to five sample requests against the analytics sandbox.
- For each, decide whether the request varies enough to need dynamic planning, or whether a fixed view would suffice.
- For any request you assign to an agent, state what governed access and verification it would need.
- Write one sentence justifying each decision.
What you produce
- A decision table (request, verdict, justification)
- At least one case correctly assigned to a fixed dashboard/workflow instead of an agent
B: Analysis Plan Critique
Critique a deliberately underspecified analysis plan and rewrite it so every required element is explicit.
Files used: Governed analytics sandbox (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Read the provided vague request and its incomplete plan draft.
- Identify which of entities, metrics, filters, time range, evidence requirements, or approval needs are missing.
- Rewrite the plan so every element is explicit and grounded in the sandbox schema.
- State what evidence the final answer must cite to be trustworthy.
What you produce
- The critiqued plan with gaps identified
- A rewritten, complete analysis plan
- A list of required evidence items
C: Tool Contract Design
Design narrow, allowlisted tool contracts for a small set of agent capabilities against the sandbox.
Files used: Governed analytics sandbox (practice, with challenges), Guided Lab starter notebook (verifiable read-only agent)
Steps
- List the capabilities an agent would need against this sandbox (for example, querying a metric summary, publishing a report).
- For each, write a tool name, an explicit parameter schema, and a read-only or write/action classification.
- Add an allowlist restricting which source systems or metrics the tool may touch.
- Mark which tools require human approval before executing.
What you produce
- A table of tool contracts (name, schema, classification, allowlist)
- An explicit note on which tools require approval
D: Semantic Metric Resolution
Given two conflicting ad hoc definitions of the same metric, resolve them into one semantic-layer definition.
Files used: Governed analytics sandbox (practice, with challenges), Data dictionary (Traditional Chinese)
Steps
- Read two different query snippets that each compute "conversion_rate" differently.
- Identify exactly where the two definitions diverge (scope, filters, or units).
- Write one semantic-layer entry: formula, unit, owner, and grain.
- State how an agent should behave if asked for a metric that is not yet in the semantic layer.
What you produce
- A comparison of the two conflicting definitions
- One resolved semantic-layer metric entry
- A stated rule for unresolvable/unknown metrics
E: Permission Boundary Test
Test whether a sample agent tool correctly enforces row/column security for a specific role.
Files used: Governed analytics sandbox (practice, with challenges), Guided Lab starter notebook (verifiable read-only agent)
Steps
- Pick a role (for example, support_analyst) and list which columns/rows it should be allowed to see.
- Run the sample query tool as that role and check whether it returns anything outside that boundary.
- If it leaks a disallowed field or row, document exactly what leaked and why.
- Propose the specific fix (a column filter, a row filter, or a deny-by-default rule) that would close the gap.
What you produce
- The role's allowed columns/rows
- A record of the boundary test result
- A concrete fix for any leak found
F: Golden Question Evaluation
Build a small golden-question set and run it against a sample agent response to check for regressions.
Files used: Governed analytics sandbox (practice, with challenges), Independent Lab starter notebook / template
Steps
- Write at least three golden questions covering a baseline case, an ambiguous request, and a schema-drift case.
- For each, state the expected evidence or expected behavior.
- Compare a sample agent answer against each expectation and mark pass/fail.
- For any failure, note whether it is a planning, tool, or semantic-layer problem.
What you produce
- A golden-question set with expected evidence/behavior
- Pass/fail results
- A root-cause note for any failure
G: Prompt Injection Defense
Identify an instruction hidden in retrieved data and verify the agent correctly treats it as untrusted content.
Files used: Governed analytics sandbox (practice, with challenges), Independent Lab starter notebook / template
Steps
- Read the provided sample document containing an embedded instruction aimed at the agent.
- Identify exactly which text is an injected instruction versus legitimate data content.
- State the correct agent behavior (refuse/ignore the embedded instruction; keep following system policy).
- Propose one detection rule that could flag this kind of embedded instruction automatically.
What you produce
- The identified injected instruction
- A statement of correct agent behavior
- One proposed detection rule
H: Verification Harness Drill
Run a sample answer through a verification harness checklist before it would be presented to a user.
Files used: Governed analytics sandbox (practice, with challenges), Guided Lab starter notebook (verifiable read-only agent)
Steps
- Take a sample answer and its underlying query result.
- Check query scope, row counts, an independent reconciliation total, and whether the claimed code actually executed.
- Decide whether the answer should be withheld, revised, or presented, and state your reasoning.
- Write the final answer with its cited dataset version and stated confidence/uncertainty.
What you produce
- A completed verification checklist
- A withhold/revise/present decision with reasoning
- A final answer citing dataset version and confidence
The done mark is stored in this browser only (and synced once you join a cohort). Keep your work in your own files or notebook.