Repository navigation
feat(reward): preserve episode outcomes and support mean-only groups - #1750
Merged
Merged
Conversation
5 of 7 tasks
dingzhiqiang
marked this pull request as ready for review
September 22, 2026 11:11
dingzhiqiang
requested review from
CormickKneey,
HwVanICI,
Le8r0nJames,
PrometheusComing,
fishcrap,
garrett4wade,
geshi001,
guozhihao-224,
nuzant,
rchardx and
sitabulaixizawaluduo
as code owners
September 22, 2026 11:11
dingzhiqiang
had a problem deploying
to
AReaL-unittests
September 22, 2026 12:00 — with 
GitHub Actions
Failure
GitHub Actions
Failure
dingzhiqiang
had a problem deploying
to
AReaL-unittests
September 22, 2026 12:00 — with
GitHub Actions
Failure
GitHub Actions
Failure
dingzhiqiang
had a problem deploying
to
AReaL-unittests
September 22, 2026 13:26 — with
GitHub Actions
Error
GitHub Actions
Error
dingzhiqiang
had a problem deploying
to
AReaL-unittests
September 22, 2026 13:26 — with
GitHub Actions
Error
GitHub Actions
Error
dingzhiqiang
had a problem deploying
to
AReaL-unittests
September 22, 2026 14:51 — with
GitHub Actions
Failure
GitHub Actions
Failure
dingzhiqiang
had a problem deploying
to
AReaL-unittests
September 22, 2026 17:17 — with
GitHub Actions
Failure
GitHub Actions
Failure
fishcrap
reviewed
Sep 23, 2026
sitabulaixizawaluduo
added
safe-to-test
sitabulaixizawaluduo
added
safe-to-test
sitabulaixizawaluduo
had a problem deploying
to
AReaL-unittests
September 24, 2026 07:29 — with
GitHub Actions
Error
GitHub Actions
Error
sitabulaixizawaluduo
had a problem deploying
to
AReaL-unittests
September 24, 2026 07:29 — with
GitHub Actions
Error
GitHub Actions
Error
sitabulaixizawaluduo
added
safe-to-test
sitabulaixizawaluduo
deployed
to
AReaL-unittests
September 24, 2026 08:57 — with
GitHub Actions
Active
GitHub Actions
Active
sitabulaixizawaluduo
deployed
to
AReaL-unittests
September 24, 2026 08:57 — with
GitHub Actions
Active
GitHub Actions
Active
7 of 15 tasks
This branch was successfully deployed
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
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.
Description
Expose mean-only group reward normalization through the existing v1 rollout/controller interfaces. Add Arena model-versus-system failure classification to retain attributable model failures as zero-reward trajectories; the receipt-protocol isolation issue noted below remains outstanding.
Missing concat-leaf rewards are not inferred from a scalar episode outcome. Explicit rewards are preserved, and reward-normalized groups with missing rewards remain rejected.
Scope and compatibility
Split from #1731 and based on latest main, including the merged runtime fixes in #1749. This does not duplicate #1746/#1748. The merge preserves main's v2 partial-group support and forwarding of min_usable_group_size. Mean-only rollout normalization remains v1-only and fails before sampling when requested with v2.
Validation
Outstanding review issue
Receipt parsing can fall through from invalid native receipts to GameAgent/legacy classification, and context-overflow handling can bypass native receipt health validation. Parser selection must follow the Harness result contract and reject invalid receipts without fallback. This reviewer finding remains open; the current merge-conflict resolution does not fix it. Unit-test success does not establish that this failure-classification path is safe to merge.