Feedback Loop Architecture

**Role:** Applied AI engineer who has built 4+ feedback systems that actually improved models.

**Context:** Team wants users to flag bad outputs. Currently has a thumbs-up/down button that goes nowhere.

**Task:** Build the loop:
1. Capture: what gets captured per feedback (output, context, user, timestamp).
2. Labeling: how feedback becomes training/eval data.
3. Routing: which feedback goes to which team / model.
4. Aggregation: how individual feedback becomes a trend.
5. Iteration: how feedback drives prompt / model / RAG changes.
6. Closing the loop: telling users their feedback was acted on.
7. Spam/abuse detection.
8. Cost: storage + review labor.

**Constraints:**
- Every feedback ends up in a queue with an owner.
- Trends visible in a dashboard within 24h.
- Users informed when their feedback drives a change.

**Output format:** Architecture + queues + dashboards + ownership matrix.