Capstone: Discover & analyze
Run discovery and analysis end-to-end on your question: grounded synthesis, the pre-committed analysis plan, the segmentation loop, and the findings memo.
Two weeks of the real work, compressed guidance edition. You know the moves; this lesson is the checklist and the traps we see capstone submissions actually hit.
Discovery week
- 1Run your 2-3 interviews with AI-drafted guides (non-leading — have AI check your questions for planted hypotheses before you walk in). Same-day synthesis, verbatim-quote rule, beliefs labeled as beliefs.
- 2Run the cross-interview pass. Contradictions and 'checkable claims' become your data checklist — exactly the Module 1 → Module 2 handoff.
- 3Update your hypothesis doc: what discovery strengthened, weakened, or added. Capstones that never update their hypothesis are either lucky or (more often) not listening.
Analysis week
- 1Get the analysis plan from AI before touching data — cuts, kill-criteria, alternative explanations. Paste it into your workspace; it's a submission artifact (reviewers check that the plan predates the findings).
- 2Run the loop: localize → find the bend → check the mechanism → test the counter-segment. Verify in your spreadsheet every number that will appear in the memo; log each one.
- 3Kill (or fail to kill) at least three alternative explanations, and record the cut that did it. 'What it isn't' is a required memo section.
- 4Write the findings memo (five-part structure) and put it through the hostile-CFO review. Fix what it catches. If the AI review finds nothing, your prompt was too polite — sharpen and rerun.
(1) The definition drifts — the memo answers a subtly different question than the scoping doc asked; re-read your signed definition before writing. (2) The hypothesis survives on loyalty — if the data is ambiguous, say so; 'inconclusive, here's what would settle it' is a passing finding and a professional one. (3) The numbers log doesn't exist — reviewers ask 'where does this figure come from?' about two numbers, chosen at random. Instant answers pass; archaeology doesn't.
Real-work data will fight you: duplicate customers, mystery columns, definition changes mid-series. Budget for it, note each issue in the memo's confidence section, and treat it as material — 'the customer counts are unreliable because the CRM has duplicates' is a finding (and, incidentally, the sales pitch for Data Foundations and Meld). Don't silently clean; visibly caveat.