1 · Tokens, never hex. Every colour resolves from docs/mockups/tokens.css. The [PROPOSED] tokens used here (--console-rail-*, --admin-*, --hub-id-*, --hub-purchase-*, the good/warn/bad triad) get added to packages/web-app/src/index.css under :root and .dark before this is coded — never inlined at the call site. Light and dark are both designed.
2 · Saved views are rows in Supabase, not localStorage. A view is configuration a coordinator relies on across devices and sessions; the column set, filter set, name and ordering are business state and persist server-side under the tenant's RLS policies. A cross-tenant view is impossible by construction, not by a WHERE clause in the query builder.
3 · The preview obeys the caller's row-level permissions. "6 of 41 matching files" is 41 files this user may see. The builder must never widen scope: adding a column or dropping a filter cannot surface a file the coordinator could not already open.
4 · Countdown colour is a summary, never the source of truth. The cd chip renders days remaining against the statutory 45- and 180-day dates held on the exchange. It is computed server-side from those dates and is read-only here — no control in this builder edits a deadline. A clock that has passed renders in the bad role with an explicit "over" label; it may never render as a satisfied or green state.
5 · Every run and every export writes an audit-log entry — actor, view definition, row count, timestamp. The footer states this, so the implementation must actually emit it. Exports carry the same NPI restrictions as the screen: no SSN, no DOB, ever.
6 · Filters are structured, not free-text. Each token is a {field, operator, value} triple validated by a Zod schema at the API boundary and compiled to a parameterised query. Unknown fields are rejected rather than passed through.
7 · Scheduling reuses the existing report schedules rather than introducing a second scheduler — the run itself is a Vectr workflow, not an in-app cron.
All exchange numbers, names and amounts on this surface are synthetic demo data.