The food-system carbon data layer: integrating automated life-cycle-assessment digital twins into urban digital twins

1 Introduction

Cities concentrate population, consumption, and emissions, and they have become the primary operational unit for climate mitigation policy. Urban digital twins, dynamic virtual representations of physical urban systems fed by sensor networks, building information models, and geospatial data, are increasingly deployed to support climate-responsive planning, building on a digital-twin concept that originated in manufacturing (Grieves and Vickers, 2017; Tao et al., 2019) and that Batty (2018) among others has since extended to the urban scale. Recent systematic reviews of the field converge on a consistent set of application clusters: building energy optimization, renewable energy integration, urban heat island mitigation, flood and hazard resilience, and mobility management. Across these reviews, one category of emissions is conspicuously missing. Food-system emissions do not appear in the application taxonomy of urban digital twins, despite the food system accounting for a larger share of global emissions on any of the boundaries discussed below, than any of the activities those twins routinely model.

Food systems contribute a substantial share of global greenhouse gas emissions, although reported estimates differ among studies. Using data from about 38,700 farms across 119 countries, Poore and Nemecek (2018) estimated that food systems accounted for about 26% of anthropogenic greenhouse gas emissions. Using the EDGAR-FOOD database, Crippa et al. (2021) estimated a higher contribution of 34% for 2015, with an uncertainty range of 25%–42%. The Intergovernmental Panel on Climate Change Special Report on Climate Change and Land (IPCC, 2019) estimated a range of 21%–37%. Although these estimates differ, much of the variation comes from differences in system boundaries and the way land-use change is considered, rather than conflicting evidence. City-level accounting bears this out. A consumption-based inventory for a major United Kingdom city put its total emissions at three times the production-based figure, driven substantially by imported food and drink (Millward-Hopkins et al., 2017), and city-scale food actions have been shown to rival interventions taken at larger scales (Boyer and Ramaswami, 2017).

At the same time, the instrumentation for measuring food-system emissions at fine resolution has advanced rapidly outside the urban-planning domain. A commercial and academic ecosystem of automated life cycle assessment platforms now produces product-level carbon footprint estimates for tens of thousands of food products, drawing on reference life cycle inventory databases, primary supplier data, and increasingly on machine learning classifiers that map products and ingredients to emission factors. Several of these platforms describe their outputs explicitly as digital twins of product supply chains, maintained dynamically so that improvements in upstream data propagate downstream. The conceptual vocabulary of these two communities, urban digital twins on one side and food-system life cycle assessment on the other, has converged on the same metaphor without the two being connected in practice.

This paper addresses that disconnect. We argue that automated food-system life cycle assessment digital twins constitute a latent carbon data layer that can be integrated into urban digital twin infrastructure, and we develop the conceptual architecture, the methodological requirements, and the governance conditions for doing so. The contribution is a named framework, the Food-System Carbon Data Layer, hereafter FS-CDL, together with a set of testable propositions about its feasibility and its limits.

We make the novelty claim carefully, and Section 2.5 reports the documented scan that located the relevant records, together with the sensitivity of that scan to retrieval depth, and rests the claim on individual screening, not on a count. Screening the intersecting records identifies work on each side of the gap and a small number of genuine near-neighbours, most closely an urban digital twin framework for forecasting neighbourhood carbon footprints from household consumption choices (Padsala et al., 2024). What that work and the other near-neighbours do not supply, and what this paper contributes, is the translation layer itself: an explicit crosswalk between life cycle stages and urban accounting categories, propagation of producer-level distributional uncertainty with attached data-quality tiers, and semantic harmonization across product vocabularies.

2 Background and related work

2.1 Urban digital twins and the absence of food

Urban digital twins have moved from visualization tools toward decision-support systems that simulate the consequences of planning interventions. Reviews published in 2021, 2024 and 2025 identify recurring archetypes, including artificial-intelligence-driven platforms, city information models and hazard-responsive twins, and map them onto applications in energy, heat, flooding and mobility (Shahat et al., 2021; Deng et al., 2021; Ghaffarian Hoseini et al., 2025). What these reviews share is a scope boundary drawn around the built fabric and its direct energy and transport flows: consumption-based emissions embodied in the goods and food flowing through a city are treated, when treated at all, as outside the twin. Deployed examples make the archetype concrete. Zurich’s twin couples a maintained three-dimensional city model to planning workflows (Schrotter and Hurzeler, 2020), and a statewide spatial twin in Victoria, Australia, has reached operational deployment across several built-environment case studies (Rajabifard et al., 2026). Each integrates buildings, terrain, mobility and infrastructure, and neither carries a food-system layer.

This boundary is defensible on tractability grounds. Building energy and traffic are observable through sensors the city often already owns. Food consumption is distributed across households, retailers, and institutions, and its emissions are overwhelmingly embodied upstream, outside the city’s administrative boundary, in agriculture and processing. The result is that urban carbon accounting has historically been strong on territorial Scope 1 and Scope 2 emissions and weak on the consumption-based Scope 3 emissions in which food dominates. The obstacle is data availability, not concept, and the food life cycle assessment community has spent a decade on that obstacle.

2.2 Automated food-system life cycle assessment and the digital twin metaphor

Life cycle assessment for food has a mature methodological foundation in ISO 14040 and 14044 (International Organization for Standardization, 2006a; International Organization for Standardization, 2006b), the GHG Protocol Product Standard (World Resources Institute and World Business Council for Sustainable Development, 2011b), and ISO 14067 for product carbon footprints (International Organization for Standardization, 2018). The empirical backbone for much of the sector is the Poore and Nemecek (2018) dataset, which produced central estimates and, just as importantly, documented extreme variability among producers of the same product. For combined beef and dairy, high-impact producers were found to emit on the order of twelve times the greenhouse gases of low-impact producers; across the dataset, impact could vary roughly fifty-fold among producers of the same item. This producer-level variance is the single most important empirical fact for the framework developed here, because any system that reduces a product to a single representative number discards information that is decision-relevant at the scale of municipal procurement.

Reference life cycle inventory databases provide the background data against which products are modeled. Agribalyse, the open database maintained by the French agency ADEME together with INRAE, covers more than two thousand five hundred agricultural and food products and underpins the French Eco-Score label. Agri-footprint, ecoinvent, and the World Food LCA Database provide complementary coverage, with Agri-footprint and ecoinvent widely used for background processes. A layer of commercial platforms sits on top of these databases and automates the assessment workflow (Balaji et al., 2025). Although commercial platforms differ in their implementation, most of them share several common features. They generally rely on large pre-computed emission-factor libraries and automatically match product bills of materials with those libraries. Most platforms also use data-quality models that distinguish primary supplier data from database averages and proxy values. In several cases, the maintained product model is described as a digital twin, a term with an established precedent in precision agriculture and farm-management research, where digital twins of crops, livestock, and farm operations have been used to couple sensor data with process models (Verdouw et al., 2021; Pylianidis et al., 2021; Nasirahmadi and Hensel, 2022). More recently, machine learning has been used to match free-text product and ingredient descriptions with appropriate emission factors. This has expanded product coverage, but it also introduces an additional source of uncertainty alongside the uncertainty already present in the underlying inventory databases (Ghoroghi et al., 2022; Romeiko et al., 2024; Martinez-Ramon et al., 2024).

Because the term is applied loosely across both literatures, and because this paper’s contribution depends on coupling two things that are both called digital twins, we adopt an explicit set of minimum criteria and apply them consistently. Following the formulation originating with Grieves and Vickers (2017) and the definitional reviews that follow it (Tao et al., 2019; Jones et al., 2020; Batty, 2018), we require four properties. A digital twin must have a specific physical referent, that is, a particular identified object, process, or system, not a class of them, which distinguishes a twin from a generic simulation model. It must be updated dynamically from observations of that referent, not a static representation. It must maintain state synchronization, meaning the digital state tracks the evolving physical state at a cadence relevant to the decisions it supports. And information must flow bidirectionally, so that the model both receives data from the referent and returns outputs that can act upon it. A representation satisfying the first three but not the fourth is more precisely a dynamic digital shadow or digital model; we use the term twin only where the fourth is at least partly satisfied, and we say which components satisfy it.

The label is applied unevenly in practice, so Table 1 classifies four named commercial platforms (Foodsteps, 2026; HowGood, 2026; Klimato, 2026; Mondra, 2026), together with the research systems that perform the same automated assignment function, against the four criteria using each provider’s public documentation. Only one describes its product as a supply-chain digital twin and states that a change in a supplier’s data propagates automatically to every affected downstream product. The others document recalculation when a user supplies new data, which meets the first criterion and at most a weak form of the second; none documents state synchronization, and none satisfies the fourth criterion upstream. An automated life cycle assessment platform is not a digital twin merely because it is automated or keeps a current database, which is the conflation Kritzinger et al. (2018) formulated their taxonomy to prevent, separating digital model, digital shadow and digital twin by the direction and automaticity of data flow. And because vendor claims are self-reported and unaudited, the table records what is publicly claimed, not what has been verified. The framework therefore requires the twin-like properties it depends upon, chiefly dynamic updating and a recorded data-quality state, to be declared and evidenced by whichever platform supplies a city’s data.

TABLE 1. Classification of named automated food-system carbon platforms against the four minimum digital twin criteria of Section 2.2.

C1, specific physical referent; C2, dynamic updating from observations of that referent; C3, state synchronization; C4, bidirectional information flow. Entries are drawn from each provider’s public documentation accessed in August 2026 and from the cited research literature. Vendor claims are self-reported and have not been independently audited, and inclusion is neither an endorsement nor a comparative assessment of accuracy. The purpose of the table is to show that the digital twin label is applied unevenly across systems performing broadly similar automated life-cycle-assessment functions, which is why explicit criteria are required. One listed provider is a former employer of the corresponding author; see the Conflict of Interest statement.

Applied honestly, the criteria classify the two coupled objects asymmetrically. An automated product-level platform has a specific referent, a product with its bill of materials and named suppliers, and it updates as supplier data, recipes and background inventories change, so it meets the first three criteria. Its claim on the fourth is genuine but narrow: outputs return to the referent through procurement and reformulation decisions and through data-quality requests to suppliers, and that channel is what the framework’s feedback path uses. Deployed urban twins generally meet the first three criteria and meet the fourth only for subsystems a municipality controls, typically energy, mobility and buildings, not the food supply chain. We therefore use digital twin for the food-system platforms in the sense defined above, while noting that the food side is closer to a digital shadow for most of the upstream chain, since a city cannot write back to a distant farm. Section 6.5 treats the consequences, and this is why the feedback path in Figure 1 returns to the automated twins alone.

Studies on eco-labels also inform the framework’s design. Potter et al. (2021) reviewed 56 experimental studies (42,768 participants, 76 interventions) and found the labelled condition outperformed the control in 60 of 76 comparisons. Three limitations bound that result: 41 of the 56 studies used hypothetical rather than real purchasing settings, where observed effects were smaller; 39 were rated at unclear or high risk of bias; and only 17 evaluated labels reporting greenhouse gas emissions specifically. Trust also moderated the effect, with labels from private commercial actors discounted relative to those of institutional provenance, which bears directly on whether a city can rely on vendor-supplied scores without an independent data-quality descriptor. Two design consequences follow. The framework prefers an interpretable scale to a single ordinal grade, because single-score labels collapse the producer-level variance Poore and Nemecek (2018) documented into a grade that cannot represent uncertainty. We treat that collapse as a design failure to be avoided rather than replicated at city scale.

2.3 Reference inventory databases: coverage, boundary, and currency

The background data that automated platforms draw on are not interchangeable, and the differences among the major reference inventories propagate directly into any city-scale figure built on top of them. Four databases dominate food-system work, and they differ on three axes that the FS-CDL must track explicitly: the system boundary each adopts by default, the geographic specificity of its unit processes, and how often it is revised.

Agribalyse, maintained by the French agency ADEME with INRAE, is built for food specifically. Its version 3 release documents more than 2,500 food products at the point of consumption, and it underpins the French Eco-Score and the product data that feed the national environmental-labelling scheme. By default, Agribalyse uses a system boundary that extends from the farm gate to the consumed product. This makes it suitable for consumer labelling. However, the reported values already include transport, processing, and some preparation stages. These need to be separated before the data can be linked with the urban crosswalk presented in Table 2. Its geographic anchor is French agriculture; using it outside France imports French production conditions unless the user substitutes local processes. ADEME revises it on a multi-year cycle, not continuously.

TABLE 2. Boundary crosswalk between life cycle assessment stages and urban carbon accounting categories.

Life cycle stage boundaries follow ISO 14040 and ISO 14067 (International Organization for Standardization, 2006a; International Organization for Standardization, 2018), and the territorial and consumption-based accounting categories follow established urban metabolism and corporate value chain accounting practice (Kennedy et al., 2010; Kennedy et al., 2011; World Resources Institute and World Business Council for Sustainable Development, 2011a). The crosswalk enforces conservation: each stage maps to exactly one category, and within-city stages are explicitly reconciled with the energy, mobility, and waste components of the urban digital twin to prevent double counting.

Ecoinvent, by contrast, is a general-purpose inventory rather than a food database. Version 3 provides on the order of twenty thousand unit processes across all sectors, with agricultural and food processes forming one part of a much larger system, and it offers a consistent set of system models, including allocation-at-the-point-of-substitution and a cut-off model, that determine how upstream burdens are assigned (Wernet et al., 2016). Its strength is background coverage and methodological transparency; its weakness for food work is that primary agricultural datasets are unevenly resolved by region and are not maintained on the food-specific update cadence that a labelling context expects. Ecoinvent releases point updates roughly annually.

Agri-footprint occupies the middle ground. It is a food- and agriculture-specific inventory, aligned with the ISO and Environmental Footprint method families, that provides both attributional and, for some processes, alternative allocation versions, which matters because allocation choice can move a product’s footprint by a wide margin for co-product-heavy systems such as dairy and oilseeds. The database mainly represents European production systems, although its international coverage has been expanding. The World Food LCA Database, developed by Quantis and its partners, provides cradle-to-farm-gate or cradle-to-processing-gate data for food and feed. It also includes methodological guidelines that define data-quality requirements and regional production mixes. Unlike Agribalyse, it is intended to be used together with a background database such as ecoinvent rather than as a standalone database.

In practice, the same product may have different published carbon footprints depending on the inventory database, allocation method, and system boundary used before any classification or primary-data adjustment is applied. These differences should not be averaged. Instead, the proposed framework keeps this information through the data-quality tiers in Table 3 and the boundary crosswalk in Table 2. This allows each city-level estimate to retain information about the inventory source, system boundary, and allocation method used. If a platform combines different inventories without recording this information, that traceability is lost, making it difficult to compare results between cities or to audit changes in a city’s emissions over time.

TABLE 3. Data-quality tiers used in the FS-CDL output, adapted from pedigree-matrix approaches to life-cycle-inventory uncertainty scoring (Ciroth et al., 2013) and informed by the growing use of machine learning classifiers to fill data gaps in life cycle assessment (Ghoroghi et al., 2022). Each line item carries its own tier. A city-scale aggregate reports its tier mixture through the coverage statistic of Section 4.5 and propagates each tier’s dispersion as in Section 5.1, so the reported interval is never narrower than its least reliable inputs allow (Wilkinson et al., 2016). Tier four carries a lower assumed dispersion than tier three because a classifier prediction targets the specific product; see Section 5.1.

2.4 Need for an integrated framework

Table 4 summarizes the two literatures along the dimensions that matter for integration. The urban digital twin literature has spatial resolution, real-time data pipelines, and a planning audience, but no food-system data layer. The food life cycle assessment literature has product-level emissions resolution and dynamic supply-chain models, but no spatial referencing to the city and no connection to urban planning workflows. The FS-CDL framework is the bridge that the comparison implies: it takes the product-resolved, dynamically maintained output of the food side and renders it spatially explicit, uncertainty-aware, and consumable by the urban side.

TABLE 4. Comparison of the urban digital twin and food-system life cycle assessment literatures along dimensions relevant to integration, drawing on the urban digital twin literature (Batty, 2018; Shahat et al., 2021; Deng et al., 2021; Ghaffarian Hoseini et al., 2025) and the automated food-system life cycle assessment and agrifood digital twin literature (Verdouw et al., 2021; Pylianidis et al., 2021; Nasirahmadi and Hensel, 2022; Poore and Nemecek, 2018). The FS-CDL framework addresses the cells in which each literature is deficient relative to the other.

2.5 What a scan of the published record can and cannot establish

The claim that the two literatures have not met is worth measuring, but the measurement must be reported with its limits. Four bibliographic databases were searched: Crossref for the registered journal record, OpenAlex for its broader coverage of indexed scholarship, arXiv for preprints, and Web of Science, whose Boolean retrieval returns a complete result set instead of a relevance-ranked sample. Nine thematic query sets spanned both poles of the prospective intersection and the ground between them, and the strings are reproduced in full in Supplementary Material S2. Duplicate records were removed, and the results were shortlisted for reading against four content criteria: digital-twin technology, food or agriculture, the urban scale, and sustainability or emissions accounting. The search was then extended by following both the references and the citing literature of every record retained after screening. The keyword criteria identified candidates for reading and nothing more; every candidate was read and assessed by the authors. The query strings, retrieval parameters, candidate vocabularies, and the inclusion and exclusion criteria applied in screening are set out in Supplementary Material S2, with the retrieval dates, since these are live indexes amended continuously.

The searches returned 6,088 unique records across the three open indexes, Crossref, OpenAlex and arXiv, once duplicates were removed, a further 2,824 unique records from the eight Boolean searches, which returned 3,805 before deduplication, and 188 records examined in the reference lists and citing literature of the retained set. Because the Boolean queries return complete sets, one count is exact rather than depth-dependent: the three-way intersection of digital-twin, food and urban terms contains 75 records for the window searched. The sources proved complementary rather than redundant, and no single one surfaced all of the records ultimately retained. Two limits on the pools are reported in Supplementary Material S2 and bound what they can establish: relevance-ranked retrieval returns more candidates the more deeply each query is taken, and optional abstract deposition leaves a substantial share of records to be assessed on their titles alone. A search of this kind is therefore a discovery device for locating candidate records, not an instrument for establishing absence, and we do not rest the novelty claim on any count.

What the search is good for is finding candidates to read, and the candidates were read. Screening at abstract level shows that most of the co-occurrence is superficial: supply-chain or inventory twin studies in which the urban tag is incidental and food is one commodity among others, agricultural and vertical-farming reviews in which the urban reference is to controlled-environment production, generic digital-twin overviews, and occasional false positives on incidental wording. Each intersecting record was screened individually, and seven have a substantive claim, in three categories. The first is urban digital twins that touch consumption or resource flows. Bonin and Frankhauser (2025) examine what it means for an urban digital twin to resemble its territory for scenario construction, but treat neither food systems nor life cycle assessment. Shehadeh et al. (2024) apply digital twins to water-energy-food-environment nexus management in an urban refugee camp, a genuine urban food-system twin application, but the food dimension is resource-flow management rather than product-level life cycle assessment, with no inventory data, uncertainty propagation or boundary crosswalk involved. Padsala et al. (2024) is the closest neighbour: an urban digital twin framework forecasting neighbourhood carbon footprints from household consumption choices, explicitly including food. It is an early-stage concept for modelling consumption-based emissions at neighbourhood resolution; it does not treat automated product-level life cycle assessment twins as a data source, address the boundary crosswalk between life cycle stages and urban accounting categories, propagate producer-level distributional uncertainty, attach data-quality tiers, or address semantic harmonization between product vocabularies. The FS-CDL is therefore complementary: their work establishes the demand case, and the translation layer specified here addresses the data-integration problem their framework leaves open. The second category is digital twins of urban food production facilities, which satisfy the criteria of Section 2.2 more fully than the product-level platforms do: a twin of a London underground hydroponic farm and a decision-support twin for urban aquaponics each observe a controllable physical referent and write back to it (Jans-Singh et al., 2020; Ghandar et al., 2021). They twin the production environment of one facility rather than the embodied emissions of what a city consumes, which is the distinction this framework turns on. The third category is city-scale life cycle assessment applied without a twin, a strand with its own terminology and research agenda (Petit-Boix et al., 2017; Loiseau et al., 2018). Dorr et al. (2022) build a multi-criteria city-scale assessment covering food, buildings and mobility and demonstrate it on a French municipality, and Elliot and Levasseur (2022) model an urban metabolism carbon footprint dynamically under dietary-change scenarios. Both establish that city-scale food assessment is feasible and policy-relevant, but neither draws on automated product-level twins nor supplies the translation layer specified here. Where the intersection is populated, it does not formalize food-system life cycle assessment twins as a structured, uncertainty-aware carbon data layer within urban digital twin infrastructure, and none of the seven reports a deployed integration in an operating municipal platform. That remains a claim about the reachable published record, not a proof of absence.

3 Hypothesis and scope conditions

We state the central claim as a two-part hypothesis, with each part testable on its own within current knowledge.

• H1 (analytical feasibility). Product-level food-system life cycle assessment digital twins can be integrated into urban digital twin infrastructure as a spatially explicit, uncertainty-aware carbon data layer using only publicly available life cycle inventory databases and open urban datasets, without requiring proprietary vendor data. H1 is a claim about the specifiability and computability of the architecture on open data. It is deliberately distinguished from the stronger claim of technical interoperability with a deployed urban digital twin platform, which this paper does not assert and which would require an implementation study to establish.
• H2 (governance dependence). The operational viability of that integration, as distinct from its analytical feasibility, is conditional on governance instruments, specifically interoperability standards for supplier data exchange, harmonized environmental footprint category rules, and reporting mandates that create demand for consumption-based urban emissions data.

H1 is a feasibility claim, so we state it as criteria that can be checked instead of judged. For a named basket of products in a named city, H1 holds if four things are true at once. Every field of Table 5 record can be filled for every line in the basket from publicly available inventory data, open-licence or licensed, and open urban data, with nothing left that only a platform vendor could supply. The Table 2 crosswalk puts each included life cycle stage in one urban category and one only, and the allocated stage emissions add back to the unallocated product total. Tier one and tier two data cover at least half the activity mass, a floor declared before the basket is run and reported in Section 5.1. The excluded-mass share of Section 4.5 stays under 10% of activity mass, a ceiling declared on the same terms. The first two criteria ask whether the architecture can be specified at all. The last two are what give H1 something to fail on. A basket can be made to look complete by descending to proxies, or by dropping the products that are hard to match. Agreement with an independently measured municipal inventory is not included, and neither is ingestion by a deployed platform. Those are the stronger claims, and we are not making them.

TABLE 5. Minimum record schema for a Food-System Carbon Data Layer exchange.

Every field is required for a conforming implementation. The schema is specified so that the framework’s four operations remain recoverable from any single record: boundary_declared and stage_set make the boundary alignment auditable; estimate_central with interval_low, interval_high, and interval_coverage carry the propagated uncertainty together with its coverage; data_quality_tier, factor_vintage, coverage_tier12_share, and excluded_mass_share carry provenance and completeness; and product_id with product_label preserves the semantic mapping. No serialization is prescribed, but an interface that omits the interval, the coverage, or the tier reduces the layer to the single aggregate figure the framework exists to avoid.

H2 concerns adoption. The three instruments are defined so that a coder working from public documents can decide whether each is present. An interoperability standard is present where a published, versioned specification for product-level footprint exchange exists and something on the public record shows it in use by someone other than the body that issued it: a conformance register, a documented implementation, a reference in regulation. Harmonized environmental footprint category rules are present where a rule set in force prescribes the calculation method for the product categories in the basket. A reporting mandate is present where a binding instrument obliges organizations in the jurisdiction to report consumption-based or value-chain emissions. We code those last two separately. A consumption-based mandate falls on the municipality, a value-chain mandate on its suppliers, and H2 expects more from the first. Presence is dated from entry into force, or for phased instruments from the date compliance first binds; voluntary and lapsed instruments are coded absent. The study is a jurisdiction-by-year panel, instruments read at year t and the outcome at t + k. One difficulty is better stated than worked around. Section 2.5 found no deployed food-system carbon layer in the reachable record, so an outcome coded on deployed layers has nothing to count. The panel is coded instead on a proxy observable now, a food or consumption-based line in the municipal inventory or a food extension to the platform’s city model, and recoded on deployed layers once enough exist. The proxy is coded present where the inventory reports a consumption-based food category as a separate line, or the schema carries a food-system extension, either verifiable from the published document. H2 weakens if the joint effect of the three instruments, with jurisdiction size and administrative capacity held as covariates, falls under a minimum set in advance, and fails if it sits inside an equivalence bound around zero. This is the test we would accept. We have not run it.

Section 5 takes H1 as far as analytical feasibility, working from published emission-factor ranges and open inventory data with stylized urban parameters. Section 6 and Table 6 set out the governance conditions the H2 panel would code. Section 7 returns to empirical validation and end-to-end interoperability.

TABLE 6. Governance conditions are mapped to enabled framework components and to the failure mode arising from each condition’s absence, reflecting documented interoperability and reporting-mandate barriers in corporate and consumption-based emissions accounting (World Resources Institute and World Business Council for Sustainable Development, 2011a; Bjorn et al., 2022), FAIR data-sharing practice (Wilkinson et al., 2016), and urban metabolism accounting (Zhang et al., 2015). The mapping operationalizes H2. Rows one, two and four are the instruments coded in Section 3 panel; FAIR practice is listed because auditability depends on it, not because it drives adoption.

Two scope conditions bound the framework. First, the framework concerns carbon and, by extension, the other impact categories carried in standard life cycle inventory data; it is not a full social or nutritional assessment, although Section 6 notes where those couplings attach. Second, the framework is boundary-explicit: it does not attempt to resolve the long-standing disagreement over whether food-system emissions are best counted at 26%, 34%, or some other figure, but instead requires that whichever boundary is chosen be declared and held consistent across the food and urban sides of the integration.

4 The food-system carbon data layer framework

The framework is organized as three layers, shown in Figure 1, and the layers differ in internal structure rather than being three parallel columns of peers. Within the source layer the components are hierarchically related: reference life cycle inventory databases, machine-learning emission classifiers, and primary supplier data exchanged under the Partnership for Carbon Transparency schema are inputs consumed by the automated food-system life cycle assessment twins, which are the single interface presented to the translation layer. Where no twin covers a product, a reference inventory record may pass directly to the translation layer, and the framework records that the estimate arrived by the shorter route. The translation layer, the FS-CDL proper, performs four sequential operations on the data in transit: boundary alignment, uncertainty propagation, semantic harmonization, and human-centered representation. The urban integration layer is likewise a chain rather than a set of parallel consumers. Harmonized estimates enter urban metabolism accounting, which supplies the urban digital twin platform; municipal planning tools and public-facing dashboards are role-conditioned views derived from that platform rather than independent recipients of the food data, and the platform exchanges state with the coupled grid, mobility, and building twins already within it. A return path carries scenario outputs and data-quality flags from the urban platform back to the automated twins alone, the only source component a municipality can write to; the twins relay whatever is actionable to their own upstream suppliers. Section 6.5 treats the consequences of that asymmetry.

4.1 Challenge one: boundary alignment

Food product carbon footprints are typically reported cradle-to-gate, covering farming, processing, and packaging up to the factory or retail gate, and sometimes cradle-to-grave, adding distribution, use, and end-of-life. Urban metabolism accounting, by contrast, partitions emissions into territorial and consumption-based categories and is sensitive to where a boundary is drawn relative to the city limit, a methodological inconsistency that reviews of the urban metabolism literature have repeatedly flagged as unresolved (Kennedy et al., 2010; Kennedy et al., 2011; Zhang et al., 2015; Beloin-Saint-Pierre et al., 2017). Integrating the two requires an explicit mapping between the life cycle assessment system boundary and the urban accounting boundary, including a rule for the within-city stages, retail refrigeration, last-mile distribution, household preparation, and food waste, that the cradle-to-gate figure omits. The FS-CDL specifies this mapping as a declared crosswalk rather than an implicit assumption; Table 2 gives the crosswalk in its general form. The principle is conservation: every life cycle stage is assigned to exactly one urban accounting category, with no double counting and no silent omission.

Retail refrigeration shows how this works. Refrigeration before the product reaches the city is an upstream cradle-to-gate burden in food LCA. The electricity a supermarket uses to refrigerate the same product inside the city is a Scope 2 emission the urban energy model has already picked up. Count it in both and it is double counted; count it in neither and it disappears. Table 2 fixes this by giving each stage a single category. Agricultural production, processing and inbound transport sit outside the city boundary as embodied Scope 3. Retail refrigeration, last-mile distribution and household preparation are territorial, and link to the grid, mobility and building twins. Food waste splits between within-city treatment and upstream embodied losses. The line between embodied and territorial runs at the city boundary and stays there.

4.2 Challenge two: uncertainty propagation

The framework retains uncertainty throughout, so that it remains available during decision-making instead of being removed at the measurement stage. Two sources are considered: inventory data, represented in reference databases as distributions rather than single-point values following the variability reported by Poore and Nemecek; and machine learning models, where an emission factor is assigned from a product’s text description and each prediction carries its own confidence. The FS-CDL carries both forward, reporting an interval and a data-quality descriptor for each aggregated city-scale estimate rather than a single value.

Figure 2 illustrates the issue. For several common food categories the published cradle-to-gate footprints vary widely between low- and high-impact producers; for beef the 10th and 90th percentiles differ by about a factor of five, with the extremes wider still. Represented by a single average, beef would hide that spread, and a municipal procurement decision could differ by the same factor depending on the supplier selected. Retaining the percentile distribution instead lets procurement officers treat sourcing as a variable, where reducing the distribution to a single score, as consumer eco-labels do, removes information that matters at the institutional level.

Each reported value has three components: a central estimate, an interval, and a data-quality tier indicating whether the result rests on primary supplier data, database averages, or classifier proxies, in the spirit of the pedigree-matrix scoring long used in life cycle inventory practice (Ciroth et al., 2013). Table 3 describes the tiers, which are adapted from the tiered approaches already used in commercial LCA platforms.

Consider a procurement example of 50 tonnes of beef per year. Published cradle-to-gate emission factors for beef vary widely. According to Poore and Nemecek (2018), the producer distribution has a 10th percentile close to 20 kg CO₂e per kg, a 90th percentile close to 105 kg CO₂e per kg, and a mean of about 57 kg CO₂e per kg. The annual footprint depends on the procurement quantity and the emission factor.

When only upper and lower bounds are available, interval arithmetic is used. If the emission factor lies in e lo , e hi and the mass m is known, the footprint lies in m × e lo , m × e hi; mass is positive, so the interval is unchanged after multiplication. For m = 50,000 kg and the bounds above: 50,000 kg × 20 kg CO₂e/kg = 1,000 tonnes CO₂e and 50,000 kg × 105 kg CO₂e/kg = 5,250 tonnes CO₂e, so the footprint interval is [1,000, 5,250] tonnes CO₂e.

However, when several independent intervals are added together, the overall result becomes the sum of the interval bounds. As a result, only the extreme values are retained, and if interpreted directly, the interval suggests a midpoint of 3,125 tonnes.

Monte Carlo propagation is used when a distribution, not just a pair of bounds, is available.

Represent the emission factor as E = exp μ + σ Z with Z ∼ N 0 , 1, and fit μ and σ to the two published percentiles. Solving ln ⁡ 20 = μ − 1.2816 σ and ln ⁡ 105 = μ + 1.2816 σ gives σ = 0.647 and μ = 3.825, so the median factor is exp 3.825 = 45.8 and the mean exp μ + σ 2 / 2 = 56.5 kg CO₂e/kg, consistent with the reported mean near 57. S1 reproduces the fit.

Drawing the footprint F = m × E over 200,000 samples yields a median of about 2,290 tonnes, a mean of about 2,830 tonnes, an 80% interval (10th–90th percentile) of [1,000, 5,250] tonnes, and a 95th percentile near 6,650 tonnes.

The two methods can now be compared. The 80% Monte Carlo interval (10th–90th percentile) is almost identical to the interval-arithmetic result because both rest on the same two percentiles, which shows that interval arithmetic reproduces the anchors and says nothing about the shape within them. Here the distribution is right-skewed: the median of 2,290 tonnes is 27% below the interval midpoint of 3,125 tonnes, while the mean is higher because a few high-impact sourcing cases extend the upper tail to about 6,650 tonnes. Using the midpoint as the central estimate would overstate the typical footprint and understate upper-tail risk. The FS-CDL therefore reports the median with the 10th–90th percentile interval and, where available, the mean and an upper-tail percentile.

Where distributional data are unavailable it falls back to interval arithmetic and flags through the data-quality tier that the values bound the stated endpoints only. Two points of interpretation avoid a common misreading. A percentile interval is not an outer bound: 20% of the distribution lies outside it by construction and the upper tail extends well beyond, so it must not be read as a worst case. Interval arithmetic does produce a bound, but only on its own premises—that inputs cannot fall outside the stated endpoints and that they move together—making it conditional on assumptions rather than a probabilistic guarantee.

The two methods also differ when many procurement lines are combined. Interval arithmetic adds the intervals directly and does not account for the partial cancellation that can occur between independent lines. As a result, the combined interval remains conservative. In contrast, Monte Carlo propagation considers both correlated and independent terms together and usually produces a narrower aggregate interval when the underlying distributions and their dependence structure are known. For this reason, the framework uses Monte Carlo propagation whenever sufficient distribution data are available. If distribution data are unavailable, interval arithmetic is retained because it provides an audit-friendly bound that can be reproduced easily.

Because the choice of parametric family is an assumption rather than an observation, it requires both justification and a sensitivity check. The lognormal family is chosen for three reasons. First, an emission factor is strictly positive, which excludes families with support on the negative half-line. Second, producer-level emission factors are strongly right-skewed: the Poore and Nemecek distributions have long upper tails driven by land-use change and low-productivity systems, and a symmetric family cannot represent a mean that exceeds the median by the observed margin. Third, the variation is generated multiplicatively, since yield, feed conversion, input intensity, and transport distance compose as products rather than sums, and a product of independent positive factors converges to a lognormal by the multiplicative central limit theorem. The lognormal is also the convention in life cycle inventory practice, where pedigree-matrix uncertainty factors are applied as geometric standard deviations to lognormally distributed exchanges (Ciroth et al., 2013; Wernet et al., 2016), so the choice keeps the framework interoperable with how reference databases already express uncertainty.

That choice was tested by refitting the same two published percentiles to two alternative right-skewed positive families, a gamma and a Weibull, and repropagating; the fitted parameters and the code are in Supplementary Material S1. Because all three families are anchored on the same percentiles, the reported 10th–90th percentile interval is almost identical across them at approximately [1,000, 5,250] tonnes CO₂e, so the informative comparison is what they do between and beyond those anchors. Propagating each through the 50-tonne line with 200,000 samples gives means of 2,830, 2,900 and 2,980 tonnes CO₂e for the lognormal, gamma and Weibull, a spread of about 5%. The medians differ considerably more, at 2,290, 2,560 and 2,750 tonnes, about 20%, and divergence is largest in the far tail, where the 99th percentiles are 10,440, 8,430 and 7,630 tonnes, about 37%. Three things follow. The mean is robust to the family choice at the precision the framework claims. The median, which the framework reports as its central estimate, is the statistic most exposed to the assumption and should be read as conditional on it. And the lognormal is the most conservative of the three in the upper tail, so the assumption does not flatter the framework’s risk estimates.

Dependence between procurement lines needs more than a single sentence, because the aggregate interval is more sensitive to the assumed correlation than to the marginal family. Emission factors across categories are not independent: they share background electricity and fuel systems, common feed and fertilizer markets and, where a city sources from one region, common climate and yield conditions. To bound the consequence, the five baseline categories of Figure 3 were repropagated under an equicorrelated structure with category-level log standard deviations, sweeping the pairwise correlation from independence to unity. The mean total is invariant at about 1,520 tonnes CO₂e, as it must be, but the 10th–90th percentile interval widens monotonically: [960, 2,210] tonnes at zero correlation, [840, 2,360] at 0.3, [750, 2,490] at 0.6 and [660, 2,640] at unity, about 1.6 times the width. Assuming independence is not conservative, so a framework that reports intervals must either estimate the correlation or state that independence was assumed; the FS-CDL requires the latter as a minimum and treats the reported interval as a lower bound on width when correlations are unknown. And interval arithmetic, which implicitly assumes perfect positive correlation by summing bounds, is the correct conservative fallback when the dependence structure is unavailable, so the earlier statement that Monte Carlo propagation yields a narrower aggregate interval holds only when the correlations are known and less than perfect.

How that structure would be identified in an operating FS-CDL is a design requirement, and the framework specifies three sources in descending order of preference. Where suppliers report under a primary-data exchange protocol, the shared background systems are explicit in the reported inventories, so correlation can be estimated directly from lines drawing on the same declared processes, grid mix or named upstream supplier. Where only reference-database entries exist, correlation can be induced structurally: background inventories in a database such as ecoinvent share named unit processes, so two footprints drawing on the same electricity, fertiliser or transport process inherit a known common component, and that shared fraction bounds the correlation from below. Where a line is populated by a machine learning classifier, predictions within a category share a model and a training set and are correlated through the model, so such lines are treated as strongly correlated within a category by default. In all three cases the assumed dependence structure must be recorded alongside the data-quality tier and reported with the interval.

4.3 Challenge three: semantic harmonization

Three specific mismatches have to be resolved. The first is product identity. A procurement record may list a proprietary product name, the supplier’s bill of materials may decompose it into ingredients, the inventory database keys its processes to its own nomenclature, and a machine-learning classifier maps free text to an emission factor through yet another vocabulary. Harmonization requires a reference ontology to which each of these is mapped, ideally reusing existing schemes, product classification codes for the procurement side, established food-composition and ingredient taxonomies for the recipe side, and the database’s own process identifiers for the inventory side, rather than inventing a fourth scheme. The mapping is many-to-one and lossy in one direction: several supplier products collapse onto one inventory process, and the framework must record that collapse as a data-quality event, because a proxy match is a tier demotion.

The second mismatch is unit and functional-unit semantics. A footprint expressed per kilogram of product is not comparable to one expressed per kilogram of protein, per serving, or per euro of spend, and procurement, planning, and public-health audiences each reason in a different functional unit. Harmonization fixes a canonical unit for storage, mass of product at a declared moisture and edible fraction, and defines the conversions to the audience units as explicit, versioned transforms rather than as ad hoc arithmetic, so that a per-serving figure on a public dashboard can always be traced back to the per-kilogram quantity and the serving definition that produced it.

The third mismatch is characterization semantics. Two figures that both claim to be carbon can rest on different global-warming-potential horizons or different treatment of biogenic and land-use-change carbon, and summing them silently is an error the boundary crosswalk cannot catch because both are nominally in the same unit. Harmonization therefore attaches to every stored quantity the characterization method, the GWP horizon, and the land-use-change convention as machine-readable metadata and refuses to aggregate quantities whose metadata conflict until an explicit reconciliation rule is supplied. This is the semantic counterpart of the conservation principle in boundary alignment: nothing is added together until the system can prove the addends are commensurable. In practice the operation is realized as a mapping service that sits in the translation layer, holds the reference ontology and the conversion and characterization rules, and emits, with every harmonized figure, the provenance chain that the FAIR requirements of Section 6 then make auditable.

4.4 Challenge four: representation across audiences

The fourth challenge is representational. The same underlying data must serve a procurement officer optimizing a tender, a planner comparing district interventions, a public health authority assessing dietary co-benefits, and a resident reading a public dashboard. These audiences differ in data literacy and in decision authority, and no single representation serves all of them. FS-CDL therefore specifies role-conditioned views over a common data model, not separate datasets. The procurement view exposes the percentile spread and the sourcing lever; the planning view aggregates to district scale and overlays the urban twin’s existing layers; the public health view couples the carbon figure to nutritional and dietary-shift indicators; and the public dashboard presents a deliberately simplified summary while retaining a link to the underlying interval and data-quality tier, so that simplification never becomes misrepresentation. It is also a governance choice, since what a public-facing view omits constrains what citizens can contest, and participatory urban twin deployments show that decisions taken at this layer set effective limits on public engagement (De Jaeger et al., 2026).

4.5 Specification of activity data, allocation, and interfaces

The architecture set out so far specifies how product-level estimates are made comparable with urban accounting conventions, but a carbon data layer is not complete until it also specifies where the quantities come from, how they are placed in space and time, what happens when they are missing, what may be disclosed, and how another system reads the result. These are stated here because omitting them would leave the framework unimplementable in practice, and because each is a place where an implementation can fail while appearing to succeed.

4.5.1 Activity data

An emission factor alone yields no city-scale quantity. Every FS-CDL output requires a paired activity term, a mass or volume of a specified product consumed or discarded within a defined population and period, and the framework treats the absence of that term as a blocking condition rather than a modelling nuisance. Four sources are realistically available to a municipality and they differ in quality. Institutional procurement records and catering management systems give itemized quantities for the food a city itself buys. Waste weighbridge and composition survey data give discarded quantities by stream. Retail point-of-sale aggregates, where a data-sharing arrangement exists, give consumption at outlet level. Household expenditure and dietary survey data give consumption at population level but require conversion from monetary or dietary units to product mass, introducing a conversion error the framework requires to be recorded as such. The first two are administrative records of the city’s own operations and can support the reporting tiers; the second two are samples or proxies and cannot, however good the emission factor attached to them.

4.5.2 Spatial allocation

Product footprints are computed on supply chains, not on places, so an explicit allocation rule is required before they enter a spatially indexed urban twin. The framework permits three rules and requires the rule in force to be declared with the output. Point allocation assigns quantities to a known facility, such as a school kitchen, hospital or retail outlet, and is the only rule supporting facility-level inference. Areal allocation distributes population-level consumption across residential zones using a declared weighting variable, typically resident population or household expenditure, and supports district comparison but not facility attribution. Aggregate treatment declines spatial allocation and reports a city total. Areal allocation is sensitive to the zoning system, so the zone definition must be recorded with the output, and district figures computed on different zonings should not be compared.

4.5.3 Temporal allocation

Background inventory data are annual averages with a vintage, whereas activity data arrive daily or weekly, and the two must be reconciled explicitly rather than by silent aggregation. The framework requires an accounting period to be declared, activity data to be summed to that period, and the vintage of every emission factor to be recorded, so that a footprint computed from a five-year-old inventory is not presented as a current measurement. Two seasonal effects are flagged rather than smoothed. Production-system seasonality means the same nominal product may carry a materially different factor when grown out of season under heated cover, the mechanism underlying the peri-urban sourcing case. Procurement seasonality means a menu’s composition, and therefore its footprint, varies within the accounting period, so an annual total computed from one period’s bill of materials is an extrapolation and is labelled as one.

4.5.4 Missing data

Coverage gaps are the normal condition of food-system inventories, and the framework handles them by a declared descent through a fixed hierarchy, not by imputation. A product is matched first to verified primary supplier data, then to a specific reference database entry, then to a category average proxy, then to a machine-learning classifier estimate; if none is available the quantity is excluded and reported as excluded, with its mass carried in a residual term so the gap is visible in the mass balance rather than absorbed into the total. Each step maps onto the data-quality tiers of Table 3, so the descent is recoverable from the output. A coverage statistic must also be published alongside every city-scale figure, namely the share of activity mass carried by tier one and tier two data.

4.5.5 Privacy and commercial confidentiality

Two distinct disclosure problems arise and they require different controls. Supplier-level product footprints are commercially sensitive, since a sufficiently detailed footprint can reveal sourcing, formulation or cost structure, and suppliers will withhold primary data if disclosure is uncontrolled, which directly degrades the tier distribution the framework depends upon. Supplier-identifiable records are therefore a restricted layer, exchanged under the contractual terms of a supplier data-sharing scheme, and published outputs must be aggregated above supplier identity. Household consumption data are personal data, and the areal allocation rule can reduce to individual disclosure in sparsely populated zones; the framework requires a minimum-count threshold per reported spatial unit, below which units are merged or suppressed, and confines household-derived outputs to district scale. Both controls are implemented through the role-conditioned views of Section 4.4. These are architectural requirements, not deployment recommendations: an implementation that publishes supplier-identifiable footprints or unsuppressed small-area household estimates is not a conforming implementation of the framework.

4.5.6 Schema and interface

To be readable by an urban digital twin the layer must expose a defined record, not a report, and Table 5 gives the minimum fields. The record is structured so that all four operations remain recoverable from it: the declared boundary and stage set make the boundary alignment auditable, the central estimate with its interval and stated coverage carry the uncertainty, the tier and the vintage carry data quality, and the product and spatial identifiers carry the harmonization. Identifier conventions are specified, because harmonization fails at the identifier layer first: product identity uses a global trade item number for retail articles and a declared catalogue identifier otherwise, and spatial identity uses the host platform’s zone identifiers, following domain extensions to standard city models that add energy and emissions attributes to an existing spatial schema instead of creating a parallel one (Agugiaro et al., 2018). The interface expectation is a read-only queryable endpoint returning these records filtered by period, spatial unit and product class, which is the minimum that lets an urban twin treat the food layer as a data source. No serialization is prescribed, but whichever is adopted must carry every field in Table 5.

5 Illustrative use cases

Four worked urban use cases illustrate the framework, built from published emission-factor ranges with stylized but realistic urban parameters. Every quantity reported in this section is an illustrative scenario result rather than an empirical estimate for any city, and the percentage changes reported for the substitution scenarios and for the boundary and data-quality sensitivities should be read as the analytical behaviour of the framework under stated assumptions. Every estimate traces back to a published emission range and carries the uncertainty information the framework requires.

5.1 Municipal food procurement decarbonization

A municipal catering system serving around 10,000 meals per day across schools and other public facilities is used as an example. The FS-CDL processes the procurement bill of materials, links each food item to its corresponding emission-factor distribution, and propagates the uncertainty to estimate the annual procurement footprint. Figure 3 compares a baseline menu with two menu-reformulation scenarios that differ in how far red-meat servings are displaced by poultry and plant-based protein items. In the baseline menu, red meat, dairy, poultry, plant-based proteins, and a residual category of other items contribute about 740, 330, 220, 90, and 140 tonnes of CO₂e, respectively, to the 1,520-tonne annual total. Substitution A represents a moderate reformulation: approximately one-third of red-meat servings (a reduction from 740 to about 480 tonnes of CO₂e) are replaced with poultry and plant-based protein dishes, accompanied by a smaller reduction in dairy-based items and a reduction in minor ancillary menu items captured under the other category, bringing the annual total to about 1,200 tonnes. Substitution B represents a deeper reformulation in which just over half of red-meat servings are displaced (a reduction to about 340 tonnes of CO₂e), with a correspondingly larger dairy reduction and a further increase in poultry and plant-based protein servings, bringing the annual total to about 1,040 tonnes. The illustrative annual footprint therefore decreases from about 1,520 tonnes of CO₂e to approximately 1,200 and 1,040 tonnes under the two scenarios, corresponding to reductions of about 21% and 32%, respectively. However, the main benefit of the framework lies as much in the uncertainty information attached to those estimates as in the reduction in the emissions figure itself.

Two sensitivity analyses are used to examine how the illustrative baseline changes under different assumptions that are not represented by a single-value approach. The first analysis considers the effect of the system boundary. The baseline footprint of 1,520 tonnes is calculated using a cradle-to-gate system boundary and follows the crosswalk presented in Table 2. Including inbound transport increases the footprint to about 1,610 tonnes (approximately 6%). When retail refrigeration and last-mile distribution are also included, the footprint rises to about 1,720 tonnes (+13%). A full cradle-to-grave boundary, including household preparation and waste disposal, increases the footprint further to about 1,950 tonnes (+28%). These percentages are illustrative and are based on declared stage shares of 6%, 7% and 15% of the cradle-to-gate total, given in S1 rather than measurements from a particular city. However, they show that two cities with similar catering systems may report noticeably different emissions if different system boundaries are applied. No single boundary is recommended over another. Instead, it requires the selected boundary to be clearly defined and reported, making differences between studies easier to identify and compare.

The second sensitivity analysis concerns data quality instead of the system boundary: as procurement data improve, the reported interval narrows while the central estimate is preserved. Because the tiers describe heterogeneous sources, from verified supplier returns to classifier estimates, they cannot be combined without an explicit probabilistic model, so the model is stated. Each tier is assigned a relative standard uncertainty ui, 8% for tier one, 30% for tier two, 60% for tier three and 45% for tier four, listed in Table 3 and of the order of magnitude implied by empirically derived pedigree-matrix uncertainty factors for inventory data (Ciroth et al., 2013). Tier four sits below tier three in assumed dispersion because a classifier prediction targets the specific product while a category average does not; the tier order ranks auditability, not dispersion alone. The error on each tier’s share is a mean-preserving multiplicative lognormal factor with log-scale dispersion sigmai = sqrt(ln(1 + ui2)), the tier contributions are summed, and the total is propagated by Monte Carlo over 200,000 samples with a fixed seed, so the result is a propagated percentile interval and not a moment approximation. Errors are taken as independent across tiers, which is an assumption and not a finding; positive correlation would widen the interval without moving the mean. On the same 10th to 90th percentile basis and against the 1,520-tonne baseline, a mostly primary profile (60% tier one, 30% tier two, 10% tier three) gives [1,310, 1,760] tonnes CO₂e; a mixed profile (20%, 40%, 25% and 15% across tiers one to four) gives [1,170, 1,920] tonnes; and a screening-grade profile with no primary data (20% tier two, 45% tier three, 35% tier four) gives [1,000, 2,140] tonnes. The profiles also fix the coverage statistic of Section 4.5, at 90%, 60% and 20%. The first two clear the 50% floor set for H1; the screening profile does not, and a basket resting on it fails H1 even though the arithmetic runs to completion. The excluded-mass criterion is not exercised here, since the stylized basket matches every line. The mean is preserved at about 1,520 tonnes throughout, while the median falls from about 1,500 to about 1,440 tonnes as dispersion grows, a property of the multiplicative model and a further reason to report the interval. The tier values and the independence assumption remain illustrative choices that a city would replace with its own. What the calculation shows is not the specific widths but that the same procurement decision rests on an evidence base about two and a half times as wide under screening data as under primary data.

5.2 Urban heat and dietary carbon co-mapping

The second use case exploits the spatial referencing that the FS-CDL adds to food data. By attaching consumption-based food emissions to residential districts through retail and household data, and overlaying them on the urban twin’s existing heat and demographic layers, the framework supports co-mapping of dietary carbon intensity against urban heat exposure and social vulnerability. The analytical value is in the overlay: districts that combine high heat exposure, high social vulnerability, and high dietary carbon intensity are candidates for integrated interventions that a single-domain twin would not surface, such as siting cooled community food hubs that simultaneously reduce heat-health risk and shift dietary patterns. No causal link between heat and diet is asserted here; it makes the joint distribution visible so that planners can reason about it. The overlay described here is an architectural capability of the framework rather than a result: no district-level co-mapping is computed in this paper, and the pattern described is the analysis the layer would enable once instantiated on a city’s own data.

5.3 Smart-city food-waste emissions monitoring

The third use case closes the boundary crosswalk of Table 2 at the disposal stage. Food waste carries a double emissions penalty: the embodied upstream emissions of the wasted product, which are lost, and the downstream treatment emissions, which occur within the city. By linking the FS-CDL to municipal waste-stream data, the framework attributes both components and identifies the waste flows whose embodied upstream burden most exceeds their treatment burden. This reframes food-waste policy from a within-city treatment problem, where most municipal attention sits, to an embodied-emissions problem, where the larger figure usually lies, and it does so using the same product-level distributions that drive the procurement use case. As with the other use cases, the attribution described is a capability of the crosswalk rather than a measured municipal waste inventory.

5.4 Peri-urban sourcing and the food-miles trap

The fourth use case tests a claim that recurs in municipal food strategy, that sourcing food closer to the city reduces its carbon footprint, and shows how the framework prevents a common error. Transport is usually a small share of a food product’s cradle-to-gate emissions, and for most products the production stage dominates, a pattern established for the United States food system by Weber and Matthews (2008) and echoed in subsequent surveys of urban food consumption footprints (Goldstein et al., 2017), so a shift to local sourcing that changes the production system, for example from open-field to heated-greenhouse cultivation to grow a crop out of its regional season, can raise total emissions even as food miles fall. The FS-CDL represents production and transport as separate, separately sourced terms rather than as a single distance-based proxy, so a local-sourcing scenario recomputes the production term for the local system rather than assuming it is unchanged.

Applied to the co-mapping of use case two, this lets a planner distinguish two superficially similar local-sourcing options: a peri-urban open-field supplier whose lower transport term is not offset by a higher production term, which reduces the footprint, from a heated-greenhouse supplier whose production term rises enough to erase the transport saving, which does not. A food-miles heuristic scores both as improvements; the framework scores only the first as one, and attaches to each the data-quality tier of the production estimate, which for a novel local supplier is often tier three or four and therefore flagged for primary-data collection before the sourcing decision is locked in. The use case does not assert that local sourcing is good or bad; it makes the production-versus-transport trade-off computable and its data quality visible, which is the distinction that the aspatial, single-figure tools in current use cannot draw. The two sourcing options contrasted here are stylized archetypes chosen to expose the production-versus-transport trade-off, not surveyed suppliers.

5.5 What the use cases establish

Each use case is instantiated on publicly available emission-factor data and open urban data structures. The use cases illustrate the analytical feasibility of H1 without testing it empirically: the architecture can be specified on open data, its four operations are computable, and they yield distinct decision-relevant outputs. They do not demonstrate technical interoperability with a deployed urban digital twin, because no such deployment is exercised here. We therefore distinguish three levels of support and place this contribution at the second: conceptual coherence, that the architecture is internally consistent and addresses the four challenges it names; analytical feasibility, that its operations are computable on open data and produce the claimed outputs, which is what Section 5 provides; and empirical validation, that an instantiation in a live urban digital twin reproduces an independently measured inventory within its stated intervals, which this paper does not provide and Section 7 identifies as the next step. No use case depends on proprietary vendor data.

6 Governance and interoperability conditions

H2 holds that analytical feasibility is necessary but not sufficient: operational viability depends on the governance conditions specified below, and specifying them is what makes H2 codeable.

6.1 Interoperability of supplier data

Product-level primary data has to move between supply-chain actors, and then into the urban layer, without a bespoke integration built at every handoff. The WBCSD-convened Partnership for Carbon Transparency publishes the PACT Methodology (WBCSD, 2025a) and the PACT Technical Specifications for product carbon footprint data exchange (WBCSD, 2025b), which define a common data model and an application programming interface for peer-to-peer exchange, so that a supplier’s verified figure can be requested and delivered in a machine-readable, method-tagged form. Under the definition of Section 3 this instrument codes present: the specification is versioned, and a public conformance register lists the solutions implementing it (WBCSD, 2026). The ISO and GHG Protocol partnership announced in 2025, with a joint working group on a product carbon footprint standard (GHG Protocol, 2025), is converging on the same object: a footprint that travels with its boundary, method, and data-quality metadata attached. For the FS-CDL, adoption of such a standard is the precondition for tier-one primary data reaching city scale at all; absent it, the city is structurally limited to database averages and proxies, the tier-two-to-tier-four band of Table 3, no matter how good any individual supplier’s own data may be.

6.2 Harmonized environmental footprint category rules

The food and urban sides must share their characterization assumptions, including the global-warming-potential horizon and the allocation rules that decide how upstream burdens are split among co-products. The named instrument here is the European Commission’s Product Environmental Footprint method (European Commission, 2021) and its product-category rules, which fix, for a given product category, the boundary, the allocation hierarchy, the impact-assessment methods, and the data-quality requirements that a compliant study must follow. Where the food side reports under a defined Product Environmental Footprint Category Rule and the urban side follows a separate inventory convention such as the GHG Protocol or a GPC-style territorial inventory, the boundary crosswalk of Table 2 must reconcile the two explicitly, mapping each life-cycle stage to its urban category under a single declared GWP horizon. Divergent characterization rules are the most likely silent source of error because the mismatched figures are nominally in the same unit; harmonized environmental footprint category rules are what allow the semantic layer of Section 4.3 to refuse an incommensurable aggregation rather than perform it unnoticed.

6.3 FAIR data principles and provenance

The inventory and classifier outputs must be findable, accessible, interoperable, and reusable, with versioning and persistent unique identifiers (Wilkinson et al., 2016), so that any city-scale figure can be traced to the database version, allocation model, and classifier that produced it. The operative instruments are the FAIR principles themselves, together with the persistent-identifier and provenance standards that implement them, including versioned dataset DOIs for inventory releases and the W3C provenance model for recording how a derived figure was computed. In the food-LCA setting this means that a released inventory carries a citable version, that a classifier model carries an identifiable version and training provenance, and that the harmonized figure emitted by the translation layer carries the chain linking it back to both. Without this machinery the data-quality tiers of Table 3 are unauditable: a tier-two figure that cannot name the database version it averaged, or a tier-four figure that cannot name the classifier that produced it, is a claim that cannot be checked, and an unverifiable tier is functionally no better than an unlabelled one.

6.4 Reporting mandates and institutional demand

Institutional requirements for consumption-based or value-chain emission reporting create a practical need to link food-system data with urban accounting, and several recent policies and standards support this direction. The EU Corporate Sustainability Reporting Directive, implemented through the European Sustainability Reporting Standards adopted in 2023 (European Commission, 2023), requires organizations within its scope to report material Scope 3 emissions, including purchased goods such as food, using defined and auditable methods. Directive (EU) 2026/470 narrowed that scope to undertakings above 1,000 employees and EUR 450 million turnover from financial year 2027 (European Union, 2026), and a revised set of standards adopted in July 2026 applies from the same year. Under the coding rule of Section 3 the mandate is present but phased, and its date of presence for a given undertaking is the first year it binds. The GHG Protocol Corporate Value Chain (Scope 3) Standard provides the accounting approach used for this purpose (World Resources Institute and World Business Council for Sustainable Development, 2011a). At city level, the Global Protocol for Community-Scale Greenhouse Gas Inventories (World Resources Institute, 2021) and the science-based-target route for cities extend the same principles to municipal reporting. Where no such requirement applies, integration remains possible but unmotivated, the failure mode Table 6 records. The design of these instruments matters as much as their existence: corporate target-setting shows a reporting mechanism can be satisfied on paper while the underlying claim is not credible (Bjorn et al., 2022), and the traceable tiers of Table 3 are meant to foreclose the food-system analogue.

6.5 Write-back asymmetry and the limits of feedback

The return path in Figure 1 is deliberately narrow, and the reason is a governance property rather than a design convenience. Of the four source components, only the automated twins are writable by a city: a municipality that discovers a mis-mapped product or an implausible interval can raise it with the platform operator, whose commercial interest is in resolving it. Reference inventory databases are third-party scientific artefacts maintained on multi-year release cycles by institutions with their own review procedures, and a city’s scenario output is not an admissible input to them. Machine-learning classifiers are typically retrained by the vendor on proprietary corpora that a municipality cannot inspect, let alone amend. Primary supplier data belongs to the supplier. The consequence is that the feedback loop improves the layer a city can actually influence, while the deeper determinants of data quality stay outside municipal reach. Cities seeking leverage there have one instrument available, which is procurement: a tender that requires declared boundaries, declared data-quality tiers, and a documented correction procedure imposes conditions on the supply chain that no feedback channel can. We therefore treat the write-back asymmetry as a scope condition on H2 rather than a defect to be engineered away, and we expect the integrations that emerge first to be those where a city already holds procurement leverage over the food flows it is trying to account for.

Table 6 maps each governance condition to the framework component it enables and to the failure mode that results from its absence. Rows one, two and four supply the instrument variables coded in Section 3 panel.

7 Discussion, limitations, and research agenda

The framework is conceptual and its claims are calibrated accordingly. Section 5.5 places the contribution at analytical feasibility. H2 is a structured conjecture; Section 3 states the panel that would test it, the condition under which it fails, and why a proxy outcome is specified.

7.1 Threats to validity

Five threats to the framework’s validity are worth naming individually, because they differ in kind and in what would resolve them.

7.1.1 The first is boundary non-comparability

Unresolved disagreement over food-system accounting boundaries is managed by requiring declaration rather than by resolving it, which means two faithful FS-CDL implementations with different declared boundaries produce different city totals. The sensitivity analysis of Section 5.1 quantifies the exposure at about 28% between cradle-to-gate and cradle-to-grave. Declaration makes the difference visible and auditable, but it does not make two differently bounded cities comparable; cross-city comparison requires a harmonization step the framework specifies but cannot enforce on its own.

7.1.2 The second is model-uncertainty opacity

The reliance on machine-learning classifiers for coverage introduces a model-uncertainty term that is harder to characterize than inventory uncertainty, because a classifier’s confidence is not a physical variance and can be poorly calibrated. Classifier output is treated conservatively as screening-grade, tier four, so that this poorly characterized term cannot masquerade as a measured one, but a conservative tier is a mitigation, not a solution, and better-calibrated classifier uncertainty is an open requirement.

7.1.3 The third is representational efficacy

The role-conditioned representation of Section 4.4 is specified as a discipline, not implemented as an interface, and the eco-label evidence of Section 2.2 warns that its effect will be design- and trust-sensitive. It is entirely possible to build a technically correct data layer whose views nonetheless fail to change decisions, and nothing in the framework guarantees otherwise; efficacy is an empirical question that only user testing can answer.

7.1.4 The fourth is data-availability bias

The framework’s outputs are only as representative as the primary data reaching them, and primary data is likeliest to exist for large, well-resourced suppliers. A city that weights its figures toward the suppliers that can report risks systematically mischaracterizing the smaller and informal supply chains that report least, which are not a random sample. The data-quality tiers record where this substitution happens but do not correct the bias it introduces.

7.1.5 The fifth is scope truncation

FS-CDL addresses carbon and the co-reported impact categories of standard inventory data; the nutritional, social, and economic dimensions of food-system policy attach at the public-health view but are not modelled here. A carbon-optimal procurement decision can be nutritionally or socially poor, and a framework that surfaces only carbon can, if used uncritically, encourage the single-axis optimization it is meant to guard against.

7.2 Research agenda

The agenda follows from the threats and separates into work achievable within roughly a twelve-month horizon and work that is inherently longer.

Short-term, within about 12 months, three studies are tractable. The first is a single-city instantiation checking the four H1 criteria on open procurement and waste data, followed by validation of the propagated intervals against an independent inventory, which is the demonstration that would move H1 from worked illustration to tested claim. The second is a benchmarking exercise running the same product basket through Agribalyse, ecoinvent, Agri-footprint and the World Food LCA Database, to quantify the inventory-choice term empirically and replace the illustrative emission-factor ranges of Section 5.1 with measured spreads. The third is a calibration study of at least one machine-learning emission-factor classifier against held-out primary data, to establish whether tier-four confidence can be made to mean something reportable.

Longer-term, three lines of work are needed and cannot be compressed. The first is the adoption panel specified in Section 3, coding the instruments of Table 6 by jurisdiction and year against the proxy outcome, and recoded on deployed integrations once enough exist to carry the test. The second is the development and controlled user testing of the role-conditioned representations of Section 4.4, drawing on the eco-label effectiveness literature to avoid the single-grade failure modes it documents. The third is the extension of the boundary crosswalk beyond carbon to nutritional and social couplings, moving the framework from a carbon data layer toward a broader food-system layer for urban planning and addressing the scope-truncation threat directly.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Google Gemini Pro was used to refine the manuscript text for language and clarity. No generative AI tool was used to create, generate, or edit any figure, and no figure content was produced by a generative model. Figures 1–3 are author-generated programmatic vector output produced with Matplotlib from the values and code in Supplementary Material S1. All AI-assisted text was reviewed and verified by the authors, who take full responsibility for the content of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fbuil.2026.1953330/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Nazish Abid, University of Bahrain, Bahrain

Reviewed by: Shun Nakayama, Senshu Daigaku - Ikuta Campus, Japan

Cumhur Dülger, Magdeburg Stendal University of Applied Sciences, Germany