The GHGI-Brain: Grounding LLM Answers in Auditable, Formula-Based Emissions Calculation

Starting in 1993, the U.S. federal government published annual greenhouse gas (GHG) inventories to satisfy international transparency commitments, but no official UN-compliant Inventory has appeared in two years. We see value in keeping the longest-running U.S. Inventory's record accessible, its methods auditable, and, crucially, its calculation approach usable by institutions without the federal government's extensive resources. Large language models (LLMs) can support all three, but closed-book LLMs fabricate numbers, and standard retrieval-augmented generation grounds prose rather than arithmetic. We present the GHGI-Brain, an auditable pipeline that retrieves from the official Inventory corpus, computes emissions with deterministic calculators we built from the Inventory's published methods, verifies that every number in an answer traces to a tool result, and cites sources the model cannot invent; the pipeline runs with a frontier API or an open 8B model on a laptop. Such a system is meant to provide continuity: it keeps an auditable national emissions record usable through institutionally tumultuous times, preserving the measurement foundation on which climate response and resilience depend. The recipe generalizes to other domains built on authoritative documents and vetted methods.