C++ Sidecar Approach Catches AI Model Label Drift in Crash Triage Pipelines

A software engineering team discovered that a free AI model endpoint used to label crash reports was silently returning incorrect triage categories despite always producing valid JSON responses. The model consistently misclassified known crash signatures — for example, labeling a heap corruption error as a network timeout — with no parsing errors or transport-layer warnings to indicate a problem. To address this, engineers built a lightweight C++ sidecar that applies a deterministic contract table of known crash signatures before treating any model output as final. When the sidecar's contract vote disagrees with the model's label, the discrepancy is logged in a drift ledger rather than triggering a retry, since retries were found to return equally confident but inconsistent answers. The approach treats the AI model's output as a proposal subject to verification, not a reliable classifier with a stable vocabulary.