Designing Label Capture

Use when the user asks how to collect eval labels from the product itself — "how do we turn thumbs up/down into eval labels?", "should we add a feedback button to the AI feature?", "can user edits, accepts, or regenerations become eval data?", "how do we get golden/labeled data without a labeling project?", "how do we capture ground truth from normal usage?", or "can an LLM pre-label our data so humans only review the uncertain ones?" — or when speccing feedback UI or a label-review queue. Walks an implicit-signal inventory, label-schema design (binary + critique, sample-size math), explicit feedback channels, zero-overhead golden-data capture, and confidence-gated agent-…