[Abhijeet Prasad (AbhiPrasad)]

🤖 This text was AI-generated.

Harbor 0.25.0 marks a trial's rewards invalid when a recorded failure makes them untrustworthy. The Braintrust plugin previously converted those rewards into normal scores:

reward=0.25, reward_valid=False -> score=0.25

The plugin now checks TrialResult.reward_valid before converting rewards. Invalid rewards produce no scores or derived metrics:

reward=0.25, reward_valid=False
  -> no scores or reward metrics
  -> raw_rewards retained in metadata.harbor
  -> warning added to metadata.harbor.warnings

Span metadata also records reward_valid and failure_cause when Harbor exposes them. Valid rewards still convert normally, including after exceptions whose policies preserve reward validity. Feature detection keeps the existing behavior on Harbor versions before 0.25.0, including the supported 0.16.0 pin.

Fixes SDK-633.
Closes #1007.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	🔄 Running since 2026-10-26T13:14:20.248288Z	177f196	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review" or "Codex (@codex) security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.