Meeting versus project knowledge, with trust light

What is said in meetings often differs from what the project knowledge says — and nobody notices. Three agents work together here: a guard protects sensitive content, an analyst finds new, contradictory and superseded statements, and a script assigns a traceable trust light from six factors. Only what passes the privacy check and carries enough trust enters the team brain. The approach is a prototype: a good 70 automated tests green, the first run on a real meeting still open. Nobody has to trust AI results blindly; their reliability can be measured on stated grounds.

Me: “Analyse this meeting against our project status.”

1. A guard agent checks the transcript for sensitive content and cleans it — or stops the run
2. An analysis agent compares statements with verified project knowledge: new, confirmed, refined, contradictory, superseded, unclear
3. A script computes a trust light from six factors: source situation, source quality, recency, retrievability, completeness, process
4. The orchestrator asks where to file it; it enters the team brain only after a passed privacy check and with a green or amber light

Assess the deltas, decide the response, approve filing.

What is said in meetings often differs from what the project knowledge says — in any organisation the contradiction surfaces late.

Any organisation with many meetings and documented project knowledge. Nobody has to trust AI results blindly; their reliability can be measured on stated grounds.