Repository navigation
Let an operator accept a live-benchmark miss with a recorded reason - #156
Merged
Merged
Conversation
The live benchmark measures one model run per case against a provider whose output varies between identical runs, so its detection rate moves several points either side of the floor without any change to the reviewer. Six consecutive versions carry a tag and no published release as a result, which means a fix for a production fault cannot reach production because an unrelated metric moved. The release workflow can now be dispatched with a reason that accepts a miss and publishes anyway. The reason is required for the release to proceed, is echoed as a warning annotation, and is written to the run summary. The baseline is untouched, so the next release faces the same floor and an accepted miss stays legible as a miss rather than becoming the new normal. With no reason supplied the behaviour is unchanged and the benchmark still blocks. This is a release valve, not a fix for the measurement. A repeated-run acceptance rule remains the way to make the floor mean something.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The live benchmark issues one model run per case against a provider whose output varies between identical runs, so detection rate moves several points either side of the floor without any change to the reviewer. Two runs of this same tree measured 91.2% and 89.5% against a 92.7% floor. The effect is not new:
v0.8.16is the last version that published assets, andv0.8.17throughv0.8.22all carry a tag and no release. A fix for a production fault currently cannot reach production because an unrelated metric moved.The release workflow accepts a
bench_live_override_reasonon dispatch. With a reason present a miss is published anyway; the reason becomes a warning annotation and a run-summary entry naming what was accepted. With no reason the behaviour is exactly as before and the benchmark still blocks.bench/baseline.jsonis never rewritten, so the next release faces the same floor and an accepted miss stays legible as a miss instead of quietly becoming the new normal.This is a release valve rather than a fix for the measurement. Making the floor mean something needs a repeated-run acceptance rule comparing a median across runs, which is the follow-up.