Payload Experimental Reverts

Experimental payload reverts automate testing by opening draft revert PRs and triggering payload validation jobs to determine if medium-confidence candidates cause blocking failures. The process runs in two phases separated by a CI wait period.

Experimental payload reverts limit operations to a maximum of 5 top medium-confidence candidates per experiment, enforcing per-experiment and cross-experiment triggering limits to control validation job execution.

Experimental payload revert results are tracked by recording per-candidate progress and results in a payload results YAML file using the payload-results-yaml skill, ensuring state persistence across both phases without creating separate tracking files.

Use experimental payload reverts when medium-confidence payload candidates require real-world CI testing to determine if they cause blocking failures, automating the creation of draft PRs and triggering validation jobs for assessment.

Experimental payload reverts integrate with payload-results-yaml and payload-autodl-json to record progress and results across phases, ensuring comprehensive tracking of draft PR validation outcomes.