Trustworthy Infrastructure Autonomy

Container upgrade workflows are largely automated: tools detect new image versions, open pull requests, and scanners enumerate vulnerabilities. The decision those workflows feed remains manual and rarely measured: given the release evidence and the deployment context, should this version be promoted to production now, or held?

This thesis makes that decision measurable. Release evidence is captured as immutable snapshots, so every evaluation is a reproducible offline replay; ground-truth promote/hold labels are produced by deterministic, policy-guided miners and adjudicated by humans; and the test split bounds training-data contamination. The method is instantiated as ReleaseGateBench, a benchmark over six widely self-hosted applications, alongside ReleaseGateCore, a reference LLM agent, and ReleaseGateStack, an open-source closed loop for production-shaped deployment.

The best instance-level accuracy is 0.703 (Claude Sonnet 5), 4.9 percentage points above the always-promote baseline, and the best macro-F1 is 0.606 (Kimi K2.6). Yet every evaluated model false-promotes roughly 68% or more of the releases the ground truth says to hold, the most accurate model getting there by promoting almost everything. The promotion decision is measurable; current agents are not yet trustworthy to make it autonomously.