Part 7 of 12 · 1 min read
The loop, and how to build it so it holds up
Five stages -- collect, triage, fix, verify, watch -- and the rules to apply.
- Never read raw thumbs as a satisfaction score; correct for response bias or collect a separate, unbiased sample.
- Treat high precision with tiny recall as seeing the loud failures only, because a detector like a downvote can be trustworthy and still miss almost everything.
- Use enough votes before declaring a winner, and keep a holdout, because winner's curse shrinks only with more votes per candidate and a confirmation sample.
- Watch for a proxy that users reward for the wrong reasons, because optimizing it will pull the product toward agreeableness instead of correctness.
- Log feedback with the trace id and version, so a signal links to the run (module 8) and to the rollout arm (module 9).
- Sample for expert labels across categories, including random non-flagged traces, to estimate recall.
- Triage into eval cases with category caps so one noisy complaint type does not fill the set (
triage_to_eval_cases). - Do not tune on thumbs alone. Require expert review and offline evals; set launch-blocking behavioural criteria.
- Guard against sycophancy: include evals where the correct answer disagrees with the user.
- Respect privacy: feedback and the conversation behind it are personal data; consent, redaction, retention (Level 10).
- Close the loop: tell users when a reported problem is fixed, if the product allows it.