{"id":"d970db23-1aa7-47d9-a718-12279393eceb","arxiv_id":"2507.19559","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A YAML-based DSL and validating parser, Sustainability Model Cards, formalizes ML sustainability reporting across training, inference, and platform dimensions, extending the Model Cards initiative.","lead":"The paper proposes Sustainability Model Cards, a new domain-specific language for describing how much energy, carbon, and water a machine learning model consumes during training and inference. It is written for teams building 'green AI' tooling, because machine-readable sustainability data would allow automatic model comparison and selection in deployment pipelines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DSL's 'precise/formal' claim is not yet supported because the stated validation rules contradict the paper's own canonical YAML example.","rationale":"The reader's weakest_assumption identifies the metamodel/validation inconsistency as load-bearing, and the same issue is the most concrete threat to the paper's central claim. The paper's contribution is a design proposal plus an initial implementation; the inconsistencies are fixable and do not invalidate the overall direction. The correct verdict remains CONDITIONAL: the DSL can be accepted as a promising first step only if the authors align the textual specification, the metamodel, and the validation rules with the worked example. My independent reading of the full text found no additional fatal flaw beyond this consistency problem; the lack of evaluation on real models is a weakness, but the more immediate blocker is that the paper's own artifact does not conform to its stated rules. A single concrete check of the parser against Listing 1 would determine whether the concern is merely typographical or reflects a genuine gap between the specification and the implementation.","tokens_in":7262,"tokens_out":2106,"duration_ms":22210,"concrete_test":"Clone the referenced parser repository (https://www.gwendal-jouneaux.fr/SustainabilityModelCards-Parser), feed Listing 1 verbatim as the YAML input, and report whether validation succeeds. Then unit-test the percentage validation rule with two minimal cards: {value: 100.0, unit: PERCENTAGE} and {value: 1.0, unit: PERCENTAGE}. If the parser accepts Listing 1 but rejects {value: 100.0, unit: PERCENTAGE}, the canonical example contradicts the stated [0,1] rule. If it accepts {value: 100.0, unit: PERCENTAGE}, the [0,1] rule is unimplemented. Either result settles whether the DSL currently delivers the claimed precision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed DSL is 'formal enough to be input of any other model description automatic process.' That claim rests on the metamodel and validation rules being internally consistent. Section 4.3 states that validation check (4) requires 'values representing percentage are bound to the [0,1] interval,' but Listing 1 uses carbon_offset_credit value 100.0 with unit PERCENTAGE and energy_mix ratio 100.0. If the parser enforces [0,1], Listing 1 is invalid even though it is presented as 'Syntax example using all the metamodel concepts.' If the parser accepts 100.0 as a percentage, then the stated validation rule is not actually implemented. Either way, the paper's specification and its canonical artifact disagree, and external tooling cannot rely on the DSL as specified. Additionally, Section 4.1 defines the root card as composed of three subcomponents (MetaData, Training, Inference), while Section 3 presents Platform as a fourth main section and Listing 1 includes a top-level platforms list; the metamodel's treatment of Platform and EnergySource is described only as a 'special case.' These inconsistencies are not cosmetic: an unambiguous schema is precisely what would make the language formal enough for automatic processing, and the current text leaves that condition unsatisfied.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Sustainability Model Cards, a domain-specific language (DSL) for describing sustainability-related information about machine learning models, including training and inference energy consumption, carbon emissions, water consumption, and platform details. The DSL is given an abstract syntax as a metamodel, a YAML-based concrete syntax with an illustrative example, and a Python validating parser generated with the authors' BESSER platform. The authors claim the DSL is 'formal enough to be input of any other model description automatic process' and outline a research roadmap for automatic model selection, deployment optimization, and SLA enforcement. The paper does not include an empirical evaluation on real models; it is a design proposal with a reference implementation.","tokens_in":7502,"tokens_out":4120,"duration_ms":36583,"significance":"If the DSL were made internally consistent, this proposal would fill a real gap: existing model reporting initiatives (Model Cards, Dataset Cards, AI Energy Score) do not provide a machine-readable, validated format for sustainability information. The paper connects quality models, Model Cards, and Green AI, and its open-source parser and metamodel are concrete artifacts that can be reused. The design choice of YAML facilitates adoption and integration with Hugging Face cards. However, the paper's contribution is currently a design rather than a validated result, and the internal inconsistencies in the specification prevent the central 'formal enough' claim from being established.","major_comments":[{"comment":"The validation rule stated in Section 4.3, item (4), requires that 'values representing percentage are bound to the [0,1] interval,' yet Listing 1, presented as a 'Syntax example using all the metamodel concepts,' contains a carbon_offset_credit value of 100.0 with unit PERCENTAGE and an energy_mix ratio of 100.0. If the parser enforces the [0,1] bound, Listing 1 is invalid; if it accepts 100.0 as a percentage, then the stated rule is not implemented. Either way, the specification and the canonical artifact contradict each other, and external tooling cannot rely on the DSL as specified.","section":"Section 4.3, Listing 1"},{"comment":"Section 3 describes the Sustainability Model Card as having four main sections—Metadata, Training, Inference, and Platform—and Listing 1 indeed contains top-level 'platforms' and 'energy_sources' keys. Section 4.1, however, states that the root SustainabilityModelCard 'is composed of three subcomponents: MetaData, Training and Inference,' and Section 4.2 calls Platform and EnergySource a 'special case.' These statements are mutually incompatible. The metamodel in Figure 1 is the authoritative abstract syntax, and the text does not state whether Platform is a component of the root class or an associated external class. This ambiguity directly affects the validation logic and any tool that interprets the schema.","section":"Section 4.1 vs Section 3 and Listing 1"},{"comment":"The paper's central claim is that the DSL is 'formal enough to be input of any other model description automatic process,' and Section 5 lists automatic model selection, deployment optimization, and SLA enforcement as scenarios. However, the paper provides no evaluation or demonstration on any real model; the only example is the synthetic YAML snippet in Listing 1, which itself conflicts with the stated validation rule. At least one end-to-end case study on a publicly available model card (with the parser run on it) is needed to substantiate the claim that the DSL is usable for automatic processing in MLOps pipelines.","section":"Abstract and Section 5"}],"minor_comments":[{"comment":"There is a typo 'aspeect' that should be 'aspect.'","section":"Section 5"},{"comment":"The conclusion refers to the 'Models Card formalism' where 'Model Cards' is the intended term.","section":"Section 6"},{"comment":"The paper inconsistently uses 'MetaData' (Section 4.1) and 'Metadata' (Section 3); please unify the spelling.","section":"Throughout"},{"comment":"The validation rule on percentage bounds should clarify whether values are expressed as fractions (0 to 1) or as percentage points (0 to 100); this would remove the apparent contradiction with Listing 1.","section":"Section 4.3"},{"comment":"The paper should state explicitly whether YAML key order is significant; the parser behavior on reordered keys is currently unspecified.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a useful design paper. It proposes a domain-specific language for sustainability reporting of ML models, with a metamodel, a YAML concrete syntax, and an open-source parser, covering training, inference, platform, water, carbon offsets, and energy mix. That fills a real gap—AI Energy Score only covers inference energy, and existing Model Cards don't handle sustainability. The survey of quality models and reporting initiatives is solid, and the writing is clear. The citation pattern is honest; self-citations to BESSER and DescribeML are relevant, not padding.\n\nThe soft spots are real but fixable. The biggest is the inconsistency the stress-test note flags: Section 4.3 says validation check (4) requires percentage values in [0,1], yet Listing 1—presented as the full example—uses 100.0 for carbon_offset_credit and energy_mix ratio. If the parser enforces [0,1], the canonical example is invalid; if it accepts 100.0, the stated rule is not implemented. Either way, the claim that the DSL is 'formal enough' for automatic processing is not yet supported. There's also a structural wobble: Section 3 lists Platform as one of four main sections, while Section 4.1 says the card has three subcomponents and Section 4.2 calls Platform and EnergySource a 'special case.' That's more than cosmetic for a schema language.\n\nThe paper has no evaluation on real models—only the one example, which conflicts with the stated rule. The reliance on BESSER to generate the parser is fine but self-referential; a second, independent implementation or a couple of realistic cards would strengthen reproducibility. None of this is fatal. The metamodel seems sensible, the scope is well chosen, and the roadmap is pragmatic.\n\nThis is a paper worth sending to a serious referee. It's early-stage design work, so I'd expect the referee to require fixing the percentage contradiction, aligning the metamodel with the text, and adding at least one realistic example before acceptance. For a workshop, it's acceptable with minor revision.","headline":"Useful DSL for sustainability model cards; the 'formal enough' claim is undercut by a validation rule that contradicts the paper's own canonical example.","tokens_in":7998,"tokens_out":2771,"would_cite":false,"duration_ms":26396,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sustainability Model Cards introduce a DSL that makes ML energy, carbon, and water reporting machine-readable and automatically processable.","keywords":["AI Models","Model Cards","Sustainability","Energy","Quality model","Domain-Specific Language","Green AI","MLOps"],"falsifier":"Run the provided parser on the paper's own example card: if a 100.0 percent carbon-offset credit is accepted despite the stated [0,1] bound for percentages, or rejected when the example intends it, the validation behavior is not as specified.","tokens_in":7074,"feed_emoji":"🌱","tokens_out":6624,"duration_ms":62938,"temperature":0.7,"pith_summary":"The paper argues that sustainability reporting for machine-learning models is still too informal for automation: current reports and energy benchmarks provide numbers but no machine-checkable structure, so sustainability information cannot be automatically compared, selected, or certified. It proposes a Domain-Specific Language (DSL) whose schema fixes the concepts a sustainability report must contain—metadata, training, inference tasks, and platform—and ties each computation to energy, carbon, and water metrics. Cards written in the language are YAML documents that a validating parser checks before they are exported in the familiar model-card format. If the language delivers the promised precision, sustainability data becomes a first-class input to model comparison, deployment optimization, and service-level agreements.","feed_headline":"AI sustainability reports become machine-checkable cards","feed_subtitle":"A new DSL covers training, inference, water, carbon, and platform data, letting MLOps tools compare and select models automatically.","key_machinery":"The mechanism that carries the argument is the metamodel of the DSL: the schema defining which concepts a Sustainability Model Card can contain and how they relate. Its load-bearing move is the Computation abstraction, from which both Training and each inference Task inherit, so that every environmentally impactful computation is described with the same energy, carbon, and water metrics plus a timestamp and a platform reference. The YAML concrete syntax maps these classes to nested key-value structures, and the validating parser checks the YAML against the metamodel before instantiating it in Python. That chain—schema, serialization, validation, instantiation—is what turns a narrative report into data.","core_discovery":"The paper's central claim is that the sustainability aspects of an ML model can be captured in a formal, machine-readable language instead of free-text or loosely structured benchmark reports. It defines Sustainability Model Cards as a YAML-based DSL with a metamodel whose root card contains metadata, training, and inference components; training and inference share a Computation abstraction that carries energy consumption, carbon emissions, water consumption, a timestamp, and a link to the platform used. A validating parser enforces unit presence and correctness, allowed inference and energy types, and value bounds, then builds an in-memory instance of the metamodel so that any other tool can process the card automatically. This makes the card simultaneously a human-readable report and a formal input to automated analysis.","pith_inferences":["Adoption of the DSL will require a shared measurement protocol; the paper defines fields and units but not how to measure, so the same model could yield different cards depending on the tool used.","The stated validation bound for percentages and the example's 100.0 percent offset credit would need reconciliation, suggesting the validation suite needs explicit tests and a decision on whether percentages are fractions or whole percents.","A concrete demonstration would be to generate cards for a handful of openly available models and run cross-model queries for deployment decisions; the paper sets this as future work but does not run it."],"forward_implications":["Sustainability cards can be validated automatically, catching malformed or unit-inconsistent reports before they enter a pipeline.","Model comparison and selection can rank candidates by environmental impact without manual reading of prose reports.","Platform details such as region, hardware, energy mix, and offsets can be reused across training and inference entries, enabling infrastructure-aware analysis.","The same formal card can be exported in the familiar model-card format for human readers and used as machine input for MLOps workflows.","Future service-level agreements can monitor and enforce sustainability commitments based on card data at runtime."],"supporting_citations":[{"why":"Supplies the motivating evidence that training large models carries large carbon costs.","marker":"[1]"},{"why":"Defines the model-reporting format that this work extends with sustainability sections.","marker":"[3]"},{"why":"Existing inference-energy rating scheme whose inference-only scope motivates the broader DSL.","marker":"[4]"},{"why":"Establishes quality assurance for AI-based systems as an open challenge that motivates formal descriptions.","marker":"[5]"},{"why":"Survey showing sustainability rarely appears in existing AI quality models, supporting the gap addressed here.","marker":"[6]"},{"why":"A dataset-description DSL that demonstrates the language-based approach this work adapts to sustainability.","marker":"[12]"},{"why":"A JSON metadata format for datasets whose interoperability goals the DSL's YAML syntax mirrors.","marker":"[13]"},{"why":"A research agenda calling for standardized sustainability metrics, which the DSL is positioned to support.","marker":"[15]"},{"why":"Low-code platform used to generate the Python metamodel implementation, backing the tool-support claim.","marker":"[18]"}],"fun_headline_variants":["Sustainability Model Cards go machine-readable","New DSL turns AI sustainability into formal cards","Formal sustainability cards for AI models","Machine-checkable sustainability cards for AI","AI energy reports become formal model cards"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim of precision rests on the language's schema and validation rules being internally consistent and exactly implemented; if the schema, the example, and the parser disagree, the cards are not yet the machine-checkable artifact described.","fun_headline_variants_meta":{"raw":{"variants":["Sustainability Model Cards go machine-readable","New DSL turns AI sustainability into formal cards","Formal sustainability cards for AI models","Machine-checkable sustainability cards for AI","AI energy reports become formal model cards"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000138,"raw_usage":{"total_tokens":1124,"prompt_tokens":888,"completion_tokens":236,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":175}},"tokens_in":504,"tokens_out":236,"duration_ms":2597,"temperature":1.0,"reasoning_tokens":175,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:00:56.159913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the provided parser on the paper's own example card: if a 100.0 percent carbon-offset credit is accepted despite the stated [0,1] bound for percentages, or rejected when the example intends it, the validation behavior is not as specified.","supporting_citations":[{"cited_title":"Mitchell, S","cited_arxiv_id":null,"evidence_quote":"Defines the model-reporting format that this work extends with sustainability sections."},{"cited_title":"URL: https://huggingface.github.io/AIEnergyScore, [Online; accessed 26","cited_arxiv_id":null,"evidence_quote":"Existing inference-energy rating scheme whose inference-only scope motivates the broader DSL."},{"cited_title":"Felderer, R","cited_arxiv_id":null,"evidence_quote":"Establishes quality assurance for AI-based systems as an open challenge that motivates formal descriptions."},{"cited_title":"Gezici, A","cited_arxiv_id":null,"evidence_quote":"Survey showing sustainability rarely appears in existing AI quality models, supporting the gap addressed here."},{"cited_title":"Giner-Miguelez, A","cited_arxiv_id":null,"evidence_quote":"A dataset-description DSL that demonstrates the language-based approach this work adapts to sustainability."},{"cited_title":"Akhtar, O","cited_arxiv_id":null,"evidence_quote":"A JSON metadata format for datasets whose interoperability goals the DSL's YAML syntax mirrors."},{"cited_title":"Alfonso, A","cited_arxiv_id":null,"evidence_quote":"Low-code platform used to generate the Python metamodel implementation, backing the tool-support claim."}],"review_version":2}