What a real AI cost-governance ledger looks like

My own routing gateway has logged 16 model calls in its current window, spending $0.356603 against a $30 cap — still a rounding error. The interesting number isn't the total, though; it's what happens to the free/paid split as real usage grows. 15 of those 16 calls resolved to a paid OpenRouter backend this window, versus 1 that stayed on a local or Google free tier — a very different ratio from an earlier capture of this same ledger, when the free tier absorbed most of the volume. The routing policy hasn't changed: free and cheap backends are still tried first, by policy, not by luck. What changed is that real call volume has grown past what those backends alone can cover, so more calls now resolve to paid credit before the $30 cap is ever at risk. 0 of 16 calls in this window settled at exactly $0 (0%) — worth naming honestly rather than only ever citing whichever window's numbers look best.

Cost governance tools that only measure spend after the fact tell you what happened. A router that decides, per call, whether a free local model, Google's free tier, or a paid OpenRouter backend actually serves the request changes what happens — the $30 monthly cap is a policy enforced before the call, with a per-request ceiling checked independently, not a number a dashboard reports after the bill arrives.

Every call records what model was requested versus what actually served it, tokens in and out, and cost — every call, not a statistical sample. That distinction matters for the same reason it matters in the governance frameworks this maps to (NIST AI RMF MANAGE-1.3, MAP, and MEASURE functions, see the framework mapping below): a sampled log can miss the one call that mattered. An unsampled one can't.