A dataset that only gets used during quarterly validation campaigns provides far less value than one wired directly into daily engineering workflow. This article covers the continuous integration architecture built around the replay harness.
Every proposed change to the V2V application layer triggers a CI job that replays a curated subset of 400 high-value traces — spanning urban canyons, highway merges, and dense intersection encounters — against the candidate build, blocking merge if precision or recall against labeled ground truth drops below the previous release's baseline.
When a replayed trace fails, the harness automatically bisects the commit range since the last passing run, re-running only the failing trace against each intermediate build to isolate the offending change within minutes rather than requiring an engineer to manually reproduce a field scenario that took a full day to originally capture.