Delivery

The proof layer

What evidence is

Not an explanation — the query that ran. Click the e# badge next to a number and you see:

• The SQL that executed
• Its parameters (site id, time window)
• The raw result rows

The numeric guard

Every number in an AI summary is compared against the values in the evidence set. A sentence containing an unmatched number is dropped — not corrected, discarded. Rounding is tolerated (the evidence may say 42.47 while the text says "42.5"); invention is not.

// a sentence the guard threw away:
// "Conversion rate jumped 41.7% to an all-time high."
//   ↳ 41.7 is not in the evidence set → sentence dropped

Causal language is locked

Correlation sentences come from templates: "coincides with the same period". The model is prevented from writing "this deploy caused that". A false causal claim destroys trust in one shot in a product like this.