# JEV-003 — Useful questions/preferences unapplied; incomplete review recovery Status: Queued; clarify product scope before implementation. PR: Pending. ## Problem and evidence The combined message “Separate decision for this browser stress test: How should we notify maintainers about critical backend errors? Slack and email are the two notification channels I want us to consider.” was held with 45% new-thread confidence despite strong question detection. A clearer follow-up created an independent alert card. “Email is the alternative alert-delivery option. It gives us a durable record, but urgent messages could be missed in a busy inbox. I prefer Slack for critical errors.” produced three “no useful role” excerpts. Review attached the email tradeoff as a Reason; the preference remained unapplied. The held-question Add-to-card form offered only Option, Reason and Constraint on existing cards, with no way to create that question. See the [live browser pass](../jev-live-browser-test.md). ## Concrete reproduction 1. Send the combined alert question/options message beside an existing monitoring decision. Inspect extraction, role and new-thread confidence/targeting. 2. Create the alert card with a clearer independent question. 3. Send the combined email tradeoff/Slack preference message; inspect held roles. 4. Use Review correction for the email reason and inspect preference recovery. 5. Review the held question and inspect creation/attachment choices. ## Expected behavior and provisional scope Useful questions, tradeoffs and preferences should apply to the intended decision or offer a deliberate recovery route. Investigate classifier/targeting policy; do not simply lower every confidence threshold. Determine whether held-question creation and preference recovery are independent product changes; if so, assign new stable issue IDs and separate PRs before implementing them. ## Acceptance checks - Record extraction/targeting evidence, distinguishing observations from policy hypotheses; cover independent questions and genuinely ambiguous messages. - Clarify intended held-question and preference recovery behavior with Maggie. - Preserve conservative targeting and prevent unrelated-card contamination. - Deterministic policy/recovery tests plus actual-browser checks as applicable. - Document and split independent recovery changes before implementation. - Separate implementer, independent tester and reviewer for each resulting PR. ## Implementation, verification and limits Queued. Classifier/targeting root cause and desired recovery scope unconfirmed.