Cryptographically Attested Environmental Accounting for LLM Inference

Jasraj Budigam

Abstract: Large language models need verifiable environmental accounting at the point of use. We present a cryptographically attested, per-request environmental accounting framework for LLM inference. The design aligns with EAT (RFC 9711) semantics and produces COSE-signed receipts that bind energy traces and environmental factors to an invocation. We embed timestamped ElectricityMaps responses in each receipt. During our runs the API returned fallback values, so we report both fallback results and a post-hoc historical correction at the same timestamps. The fallback contrast between GB and continental zones produces an apparent 7.5$\times$ spread. Historical correction changes regional ranks and lowers mean per-request CO$_2$ by about 32\%. Because clipping creates left-censoring, we report Tobit 0.162 J/token, Hurdle 0.198 J/token for non-zero requests, and OLS 0.140 J/token on all requests. Across 384 requests we observe a mean of 2.618 J per request and a mean ratio of 0.0215 J per token, while the regression slope captures marginal energy per token. Verification runs in 0.109 ms and signing in 1.066 ms, negligible relative to inference