test(eve): clarify stale-ask-question-selection judge criterion (both variants)#923
Merged
Conversation
The eval's closedQA judge flipped a valid reply to 0 ~22% of runs. Root cause
is the judge, not the model: instrumented repro showed 16/16 replies correctly
engage with the stale selection ('that response is stale and no longer
relevant; the earlier prompt was already answered'). The old criterion asked
the binary grader to confirm the reply 'acknowledges a late selection of the
STALE-CANDIDATE-7Q4M option' — the option-specific phrasing made it
intermittently answer 'no' for replies that engage in general terms.
Reword the criterion to match how models reliably respond (treat it as a late
response to an EARLIER question; apply or explain it's already-answered/stale;
'yes' unless the reply answers the current pending question or ignores it).
Verified against a deployed preview: 0/15 flips post-change vs ~22% before.
Signed-off-by: prha <prha@vercel.com>
prha
marked this pull request as draft
July 17, 2026 22:28
Contributor
|
The latest updates on your projects. Learn more about Vercel for GitHub.
2 Skipped Deployments
|
prha
marked this pull request as ready for review
July 17, 2026 22:53
The -idle sibling carries the identical closedQA criterion and the identical flake — run 29618944115/attempt-4 failed on BOTH stale-ask-question-selection and -idle scoring 0%, and the earlier commit only fixed the non-idle one. Apply the same clarification, adapted for the idle flow (nothing is pending when the stale selection arrives, so no 'current pending question' clause). Verified against a deployed preview: 11/11 closedQA=100% post-change. Signed-off-by: prha <prha@vercel.com>
prha
changed the title
AndrewBarba
approved these changes
Jul 18, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
hitl/stale-ask-question-selection'sclosedQAjudge flipped a valid reply to 0 on ~22% of runs. The root cause is the judge, not the model.Evidence
Instrumented repro (16 runs against a deployed preview) showed 16/16 replies correctly engage with the stale selection — e.g. "That response is stale and no longer relevant; the earlier prompt was already answered with 'Use current context.'" Every reply satisfies the intended criterion, yet the real eval still scored 0% ~22% of the time. So the binary
closedQAgrader was intermittently answering "no" for good replies.The likely trigger: the old criterion asked the grader to confirm the reply "acknowledges a late selection of the STALE-CANDIDATE-7Q4M option" — the option-specific phrasing made it flip for replies that engage in general terms ("stale / earlier prompt / already answered") without repeating the option text.
Fix
Reword the criterion to match how models reliably respond: treat the message as a late response to an earlier question; apply it or explain it's already-answered/stale; answer "yes" unless the reply instead answers the current pending question or ignores it.
Verified against a deployed preview: 0/15 flips post-change vs ~22% before. The deterministic HITL mechanics (stale selection becomes a new user turn, fresh request ID, etc.) already pass and are unchanged — the judge was the only nondeterministic gate. Model behavior was never the problem, so no prompt/instruction change.