Worksheets

Guess the corpus size and you either ship bugs it never reached, or burn compliance hours on cases the feature never touches. Compute a defensible floor from your feature's code paths and edge cases, not a guess.

Engineering hours are the visible cost of building synthetic data in-house; compliance review and regulator-tracking refreshes are what teams forget to count. Compute the real cost both ways and the payback period.

A blank cell on this grid is a bug waiting to happen: code paths down the rows, demographic and account slices across the columns. Fill it in and see which intersections your corpus quietly skips.

Comparative rationale is the most-cited Care Obligation gap — the file has the suitability fields but never the comparison to the alternative. List every recommendation type and capture that evidence first.

A monitoring rule without a test case is a rule you can't defend in an exam. Map each typology you monitor — structuring, layering, source-of-funds, PEP-adjacent — to the synthetic transactions that fire it.

Subprocessors are the most-missed line item here — the vendor's vendor handles your NPI too. Build the system-by-system inventory the Safeguards Rule risk assessment requires, mapped to each 16 CFR 314 control family.

A TLH engine that ships without exercising every wash-sale variant ships with a footnote — cross-account cases, especially spouse-IRA repurchases, are what most teams miss. Design those cases before they ship broken.

Handling the median client well is unremarkable; what makes a robo-advisor defensible is how it responds to retirement, divorce, and panic-sell signals. Map the data signal and response for each event.

"Test the agent's memory" isn't a test plan until it's broken into five concrete question shapes — recall, rationale, precedent, ordering, attribution — each with a real scenario and a known correct answer to check against.