Decide What the AI Should Trust

Definitions tell the agent what a value means. They do not tell it what to do when two well-defined fields disagree. Acme still has that problem: a nightly risk score says high risk while a manual health field says Green.

Authority comes before convenience

Start with provenance. Who or what writes the field? Then check freshness, update cadence, and intended use. Acme’s score is recalculated nightly from product and support signals. Its health field was last touched 143 days ago by a process the team stopped using.

The score is authoritative for renewal risk. The health field can remain visible as supporting history, but it cannot clear the account.

Use a ladder the model cannot rearrange

Platform permissions sit at the top. Hard controls at the tool boundary come next. Approved context rules follow, then field help text, and finally inference from labels. Specificity breaks ties only inside the same authority tier. A narrow field rule never overrides a platform permission.

Conflicts need a real state

Two active rules can genuinely contradict each other. Do not let retrieval order pick a winner. Mark the result conflicting, stop the run, and send the pair to review. A visible error is cheaper than a number somebody acts on for three months.

In the GTM Lab, use provenance and freshness to choose the authoritative Acme signal.

Reference

• Context precedence and conflict handling →
• AI-generated fields and compounding fabrication →