ESGlass: Glass-Box ESG and Sustainability Reports

We introduce ESGlass, a glass-box paradigm for ESG and sustainability reports. Instead of treating the report page, file, or tagged fact as the native unit, ESGlass treats each material disclosure claim as an interactive evidence object that binds multimodal observations, derived computations, provenance, uncertainty, and stakeholder-specific renderings. This shift matters because sustainability evidence is increasingly dispersed across invoices, tables, sensor streams, forms, maps, satellite imagery, facility video, and in ternal workpapers, while generative AI makes polished but weakly supported narrative cheap to produce. Building on progress in ESG benchmarks, sustainability knowledge infrastructure, document AI, multimodal retrieval, agents, geospatial foundation models, and provenance standards, we formalize the report as a policy conditioned view over claim–evidence–provenance graphs and distinguish asset provenance from claim provenance. We argue that glass-box reporting demands stronger targets than citation-style grounding, including minimal sufficient evidence sets, challenge retrieval, replayable transformations, omission semantics, and selective abstention. We outline a one-company energy-disclosure prototype, define task families and evaluation criteria, and surface governance issues such as selective disclosure, privacy-preserving drill-down, and false completeness. ESGlass reframes ESG and sustainability reports from polished narrative artifacts into inspectable multimedia disclosure objects, offering a concrete multimedia research agenda for systems that must not only generate disclosure, but also defend it.