Quality-aware automation for LCI database mapping

Abstract

Scope 3 Category 1 emissions (purchased goods and services) consistently dominate corporate carbon footprints yet remain among the hardest categories to measure precisely. Activity-based accounting requires mapping each procurement line item to a lifecycle inventory (LCI) database activity, but the exact product rarely exists in the database. Every mapping is therefore a proxy selection under incomplete information, and proxy errors are silent: unlike misclassification in standard supervised tasks, an incorrect mapping produces no anomalous signal, making errors undetectable without item-level expert review. We present a two-agent system that separates proxy selection from quality assessment through an information barrier. A mapper proposes LCI database matches via iterative, tool-augmented retrieval; an independent judge, seeing only the input, the proposed activity, and its metadata, scores each mapping along two dimensions (emissions similarity and material similarity). On 1039 items spanning seven product categories, the mapper achieves 90.7% defensible accuracy (the share of items mapped to an expert-approved option) with zero abstentions, compared to 19–43% for retrieval baselines and prior systems that abstain on 70–73% of items. At a 20% review budget, the judge captures 67% of all errors versus 37–40% for heuristic baselines, and 74% of severe errors (>\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${>}$$\end{document}100% relative emissions deviation). The information barrier produces calibrated quality scores that support auto-accepting 30% of mappings at 0.3% error rate. This enables organizations to move from category-average estimates to item-level activity-based footprints while concentrating expert review where it matters most.