The decision_log is only as useful as the questions you ask it. This week, the five queries operators reach for in their first week — the ones that turn an append-only table into an operating dashboard.
The five
- refund-rate-by-rollback — does reversing an action correlate with fewer refunds downstream?
- judge-floor-misses-by-channel — which channel generates the most rejected proposals, and why?
- tier-A-conversion-uplift — how much does verified-cost confidence improve action outcomes?
- reversal-latency — how long between an action and its rollback, by agent?
- evidence-recurrence — which evidence rows show up across the most decisions?
None of these are exotic. They're the questions a good operator already asks — the difference is that with an append-only log, the answers are queryable facts instead of reconstructed guesses.