Decision 04 of 05 · 2 min read
The backend authors the screen
Why can a reader trust what the page shows?
The backend turns each pipeline result into an A2UI surface: a declarative description of a status, a chart, a table, a month picker and an evidence panel. The browser renders only an allowlisted set of those components. The model never generates the screen.
Because the screen is data, it can be stored, compared and replayed. A reloaded page rebuilds every answer from the server's history without running anything again, and choosing another month is an action on a stored answer that the server checks before it runs. The replay site uses the same renderer on recorded answers.
The trade-off
The screen is only as flexible as the component catalog. Drill-down and filters beyond dates are deliberately not built: they need a reviewed extension of the metric's contract and new expected answers first.
What was shown
- StagingFor every run checked, the rows, SQL, checks and source on screen matched the stored answer exactly.
- StagingWith browser storage cleared and the page reloaded, nine recorded turns were rebuilt from server history and no run was added.
- StagingEvery attempt to read, cancel, submit to, stream or act on another person's conversation was refused, with no run and no spend.