The Brain / Capabilities / Claim check
PROVEWhat it tells you is checked
against what happened.
Before a reply reaches you, WorkElate Chief, the Brain, compares it to what the systems actually returned, never to what you asked for.
An update that matched nothing, narrated as done
Success reported, nothing modified
The filter used an account identifier superseded by a merge. The store answers: operation succeeded, zero records changed.
The draft says the date moved
The reasoning step saw a success flag and a request to move a date, and wrote the sentence they imply.
The sentence loses to the data
The payload carries a modified count of zero, so the reply says the update matched no records and names the identifier.
A cheap judge, a narrow question, an exit
Payloads cannot be persuaded
The evidence is the structured result of the calls that ran. Text the AI read earlier, including inbound mail, is not evidence.
Abstaining is a first-class answer
The judge answers in a fixed shape and may say it cannot tell, so it never invents a verdict on ambiguous evidence.
One re-draft, then a plain correction
The AI gets one attempt holding the data. If that still misses, code writes a plain correction. Nothing ships silently.
Three honest limits
It does not grade whether the action was wise.
It runs only when something changed and completion is claimed.
A judge can be wrong; it errs towards letting truth through.
Updated 15 September 2026 · WE_AINA