Generated output should be checkable: text cites only what ran, every value says where it came from, and a query ships only with proof.

The problem

A generated dashboard looks just as confident when its numbers are made up as when they came from real data. The model writes “Review the 12 late orders” beautifully, and attaches a query that never ran. “No results” turns out to mean the filters were loosened until the question changed. After one wrong number, people stop trusting all of it.

A smarter model doesn’t fix that. What fixes it is making each of those failures impossible to ship quietly.

What I concluded

The model keeps the creative work: phrasing, choosing, laying out. Code decides what counts as true.

Cite, don’t author

The writing step may only reference queries and ids that the evidence step actually produced. Matching is exact. Anything else is stripped, and nothing fails: the sentence survives without its button.

const known = new Set(findings.map((f) => f.query));

const actions = parsed.actions.map((a) => ({
  ...a,
  query: known.has(a.query) ? a.query : undefined,
}));

The prompt asks the model to copy queries exactly. The filter is what makes it true. Every link a person can click goes to something that ran.

Every value carries its origin

Each generated value is tagged with where it came from, and the UI shows it:

  Open orders     1,284   real      ← query, as of 9:14
  Summary         "Most…" authored  ← AI-written badge
  Delivery time   (none)  gap       ← would need a shipping integration
  Trend           ~~~     mocked    ← sample, not real

Only a producer that actually read data may say real. Anything else gets demoted. A gap is an answer: “you’d need X” beats a confident zero. The origin is part of the accessible name too, so a screen reader hears the same honesty.

Generation you can verify

  • Born in the consumer’s shape. Whatever renders the output publishes its shape, and that goes into generation as a constraint. Nothing renames columns afterwards.
  • Empty needs proof. A query leaves only if it returned rows or a clean run confirmed zero. Otherwise it gets one repair turn: diagnose, never loosen. “Couldn’t check” is not a real zero.
  • Plan once, read live. A generated page stores its plan (cards and queries), never its numbers. Every view runs the queries fresh.
  • A plain floor under every model call. A banner is built from what’s in hand first. The nicer wording replaces it only if it arrives in time.
  • Personalize the planner, not the generator. What someone works on picks which questions to ask. The query for a question is the same for everyone.

Rules

  • Exact citation or no link. Rejecting a harmless reformat is the cheap side to fail on.
  • Only grounded producers say real. Everything else is authored, gap or mocked.
  • Show gaps, never zeros. Missing data says what it would need.
  • No proof, no ship. Rows or a confirmed empty, nothing else.
  • Store plans, not values. Numbers are read at view time.

What you get

Trust becomes something a person can check value by value instead of all or nothing. Each quiet failure gets a place where it’s caught: the shape at generation, the proof before return, staleness at read. The model is no longer the last line of defense.

Related: Knowledge store, Empty states, Query reuse