Govern a Foundry multi-agent code review

Why it matters

A recommendation that can delay a release needs evidence a reviewer can inspect, and this exercise builds that evidence without retaining the sensitive inputs behind it.

What to practise

Deploy a Foundry project, a model with Foundry’s default guardrails and Application Insights with Bicep, then complete a deterministic policy gate and run a security reviewer, a fairness auditor and a governance orchestrator. Check in the Foundry traces that only the security reviewer receives source code. Change one fairness probe and watch the disparity and the human-review decision change. Query the evidence with KQL, and confirm a privacy query returns no rows.

How it fits

This is 1.1.9’s accountability made concrete: response IDs keep attribution, and the review gate stays in code rather than with the orchestrator, as 1.1.6 places approval. Each handoff carries only the data its role needs — compare it with 1.1.3’s delegation contracts.