Embedded AI assistant: identity and screen-context acceptance worksheet Editorial worksheet, not a product configuration/import or security certification. Owner / review date / deployed version: Application login/session verifier: Bot identifier and permitted database/reporting surface: Trusted identity fields and their source: User-selectable page/time/segment fields and validation: Policy variables (names, types, allowed values; NO credentials): Database-side permissions and effective execution identity: Context persistence scope: request / customer / conversation / tab: Role revocation and logout mechanism, measured enforcement delay: Record EXPECTED / ACTUAL / evidence reference / owner for each: 1. User A raw rows and paid aggregate, expected A1+A2=12500 cents. 2. User B same question, expected B1=99000 cents. 3. Browser supplies B's accountId while signed in as A: reject. 4. Browser supplies role=admin or databaseConfig: reject. 5. Unknown user, missing identity or invalid policy value: reject. 6. New page/date selection changes answer scope, not permitted account. 7. Ask for all accounts, joins, exports and follow-up: stays authorized. 8. Two tabs select different periods: each answer uses its declared scope. 9. Account switch, role downgrade, logout: old iframe/history/downloads reviewed. 10. Empty authorized period is distinguished from failed authorization. 11. Saved answer/chart states period, units, filters and context revision. 12. Mock-context debug result repeated through actual authenticated user path. The downloadable Python example verifies only its local mapping and reference SQL. Unrun checks remain UNVERIFIED. Never paste session URLs, tokens or private rows here.