Consequential metrics are commonly described as instruments for representing performance, quality, or progress. This article argues that, once embedded in institutional routines and linked to consequential decisions, metrics are better understood as epistemic infrastructures: durable arrangements of categories, proxies, data systems, interfaces, incentives, norms, and revision authorities through which value is made visible, comparable, and governable. The article asks under what conditions metrics shift from provisional proxies to infrastructures, and what epistemic and governance consequences follow from that shift. Drawing on infrastructure studies, the sociology of quantification, audit studies, global indicator research, and theories of performativity, the article develops a mid-range conceptual framework for this transition. It distinguishes metrics, metric regimes, and metric infrastructures; reconstructs a layered metric infrastructure stack; and identifies a capture process in which proxies, interfaces, incentives, adaptation, and institutional embedding become mutually reinforcing. Comparative discussion of machine-learning benchmarks, citation-based research assessment, and ESG ratings shows that metric infrastructures can coordinate inquiry and improve accountability while also producing proxy drift, false commensuration, epistemic narrowing, governance asymmetry, and lock-in. The article therefore treats metrics neither as neutral representations nor as intrinsically corrosive instruments, but as socio-technical arrangements whose epistemic value depends on how they are designed, governed, revised, and contested.