CARA framework cuts cloud consolidation costs with 3.57% reliability gain

Cloud consolidation typically relies on simulator-based placement pipelines that test many repair actions but only deploy when multiple reliability criteria improve simultaneously. The problem: reusing search scenes to validate the chosen action invalidates the evidence, and scalarizing criteria can sacrifice the weakest one. CARA (Certificate-Aligned Recomposition) fixes this with an incumbent-anchored approach that cleanly separates adaptive proposal generation from a one-use deployment decision.

Within a bounded two-host neighborhood, CARA uses an admissible bound to recover the exact top-P distinct actions under a mixed paired-binary/fixed-bet certificate order. It then commits one action's betting plans before running fresh paired certification on held-out views. The authors prove the probability of falsely declaring four-way improvement is at most .05, regardless of upstream search complexity. In a prospective frozen study over 128 independent synthetic environments with four repeated contexts, the fail-closed terminal policy improved a development-selected same-host-count incumbent on all five evaluation endpoints in every environment. J dropped by 3.57 percentage points, continuous burdens by 21–28%, and matched ordering sensitivity showed absolute mean gaps below 5.3×10⁻⁴. CARA couples exact, auditable proposal recovery with selection-robust fresh certification for dependable last-mile local repair.