Large language models are good at reasoning about structure and producing well-organized drafts from structured inputs. They are not reliable emission factor databases. A model trained on general internet data has ingested thousands of pages of sustainability reports, academic papers, and policy documents that reference emission factors. But those factors span multiple countries, multiple years, and multiple methodologies. A factor from a 2019 EPA table is not the same as the 2024 version. A UK DEFRA grid electricity factor does not apply to a US facility. A CO2-only factor applied where a CO2-equivalent factor is required will undercount the footprint.
The failure mode is subtle. If you ask an LLM to estimate the emissions from a volume of natural gas and it produces a number, that number might look plausible. The factor it used might even correspond to something from a real table. But without knowing the specific source, the vintage year, and the gas type coverage, you cannot verify it, cannot cite it in a regulatory disclosure, and cannot defend it to a third-party assurer. A plausible number without a verifiable source is precisely what rigorous sustainability reporting is trying to move away from.
How Verdant grounds every figure
Verdant does not ask a language model to recall emission factors from its training data. It maintains a structured factor library, versioned and attributed to specific sources, that is separate from the LLM inference step. When a calculation is needed, Verdant retrieves the factor from the library, performs the arithmetic in deterministic code, and passes the full calculation to the drafting model as structured context. The model drafts the prose and methodology description. The number and the factor come from the library, not from the model's weights.
The factor library is organized by source, geographic scope, gas coverage, and vintage year. Before a factor is applied to a calculation, the retrieval step confirms that the factor matches the reporting year, the jurisdiction, and the methodology the customer has selected (location-based or market-based for Scope 2, for instance). If no confident match is found, the system surfaces a gap for the user to resolve rather than falling back silently to an approximation.
What the output looks like
Every emissions line in a Verdant-drafted report shows three things alongside the figure:
- The activity data: the raw input from the customer's operational records, such as the meter reading, the fuel volume in the invoice, the flight distance from the booking system, or the spend amount from the accounts payable export.
- The emission factor: the specific factor applied, with the source name, table reference, and vintage year cited. For example: US EPA GHG Emission Factors Hub, February 2024, Table 1, natural gas, combined CO2-equivalent combustion factor.
- The arithmetic: the activity quantity multiplied by the emission factor, written out explicitly, so any reviewer can verify the multiplication independently without needing to trust the system's output.
That format makes the report auditable at the individual figure level. A third-party assurer can check the factor against the named source, verify the activity data against the customer's internal records, and confirm the arithmetic. The calculation detail view exists specifically to support that verification step.
Confidence labels and where they appear
Not all emissions figures carry the same precision. A meter reading from a building management system connected directly to the platform carries higher confidence than a spend-based estimate for a procurement category where supplier data is unavailable. Verdant applies a three-tier confidence label to each figure:
- High: direct measurement or metered data with a supplier-verified emission factor. The calculation inputs are directly traceable to primary sources.
- Medium: estimated activity data with a recognized average-data factor (such as DEFRA average-data supply chain factors), or direct measurement with an EEIO sector factor. One input is an estimate or an average rather than a primary measurement.
- Low: spend-based estimate with an EEIO sector factor, or estimated activity data with a proxy factor applied to an undisclosed or partially disclosed category. Both major inputs carry meaningful uncertainty.
These labels travel with the figure through the draft report. They do not disappear when the document is formatted for publication. A sustainability lead reviewing the draft can see at a glance which figures are well-grounded and which should be prioritized for improvement in the next reporting cycle. That visibility is part of what makes the output useful for planning, not just for disclosure.
Model-agnostic means the grounding holds regardless of the model
Verdant routes drafting tasks across language models depending on availability, performance, and the nature of the task. The emission factor retrieval and arithmetic happen before and outside the model call, in deterministic code. That means the grounding discipline is not dependent on which model is active at any given time. Whether the drafting step runs on a model from one provider or another, the emission factor came from the library, the arithmetic was computed before the model saw the calculation, and the output includes the full sourced calculation.
The model's role in this pipeline is to draft the methodology description, organize the section structure, and produce readable prose that accurately represents the grounded calculation. It does not choose the factor, perform the arithmetic, or assign the confidence label. Those are deterministic operations that run before the LLM is invoked and are not subject to the LLM's tendencies toward confident-sounding approximation.
A plausible number without a verifiable source is exactly what sustainability reporting is trying to move away from. We solve that before the model sees the calculation.
The result is a draft where the creative work (prose, structure, reasoning about what to include and how to frame it) is assisted by a language model, and the numerical work (factors, arithmetic, confidence labeling) is deterministic and source-linked. Producing output that is both well-written and traceable is a harder target than either quality alone, and the separation of concerns between the model step and the calculation step is what makes it achievable in practice.
What to watch for when reviewing AI-assisted sustainability reports
If you are reviewing a sustainability report drafted with AI assistance, from Verdant or any other tool, the questions to ask before accepting any figure are the same questions a third-party assurer will ask. Where did the activity data come from and how was it collected? Which emission factor was applied and from which specific source and vintage? Does the arithmetic produce the stated result? If the tool cannot answer those questions in the output itself, the figure needs to be verified or replaced before the report is published.