Where Feedback Enters: An Artifact-Grounded Audit of a Production Meta-Agent Loop

Production meta-agents are often described as closed feedback loops, but recording an outcome does not grant it authority to alter future governance. We audit a deployed meta-agent that proposes, reviews, merges, and monitors changes to a live software product. From 676 proposals, 2,500 trace events, dated policy artifacts, 419 deployment labels, and 377 usage labels, we reconstruct a policy-feedback topology: a mapping from signals to the decisions each was permitted to influence, distinguishing implemented routes, observed activation, and traceable consequence. Review verdicts directly governed admission; deployment outcomes were eligible to raise, but not lower, a future admission threshold; usage entered proposal generation only as weak, non-vetoing evidence; and rejected alternatives rarely received downstream labels. Instrumentation expanded across six artifact-reconstructed policy eras, while evidence of realized governance effects remained sparse. The case shows why auditability does not imply evaluability: responsible meta-agent governance must version not only action policies, but also which observations may update which decisions.