Public concern about farm animal welfare is increasing (European Commission, 2023). A growing public demand for higher welfare standards puts pressure on policy-makers, farmers, and scientists to establish more welfare-friendly housing systems (Alonso et al., 2020; European Commission, 2022; Keeling et al., 2022). Concern for farm animal welfare is part of the increasingly urgent need to develop food systems and societies that are sustainable overall, environmentally and socially, as well as economically. To support the societal demand for a more sustainable food system, a valid, reliable, and transparent evaluation procedure is called for to facilitate the assessment of livestock welfare in different housing systems based on the available scientific information. Direct “measurement(s)” of animal-based features have been used by scientists (e.g. qualitative behavioural assessment (QBA, e.g. Cooper and Wemelsfelder, 2020), judgement/attention bias tests (Jones et al., 2018), and Welfare Quality^® (Blokhuis et al., 2010)), with their respective strengths and weaknesses. Nevertheless, a science-based overall farm animal welfare assessment remains a major challenge that, in the past, has been addressed mainly using expert surveys and panels focussing on expert consensus (Anonymous, 2001; EFSA, 2007, 2010, 2012, 2015, 2023; Bracke et al., 2008a, b). A more formalised scientific method that makes more explicit use of the numerous studies that have been conducted over the past decades would be beneficial to assess overall welfare (Heinonen et al., 2013; Blanco-Penedo et al., 2020; Verdon, 2022). Besides this, the rapid developments in artificial intelligence (large language models) may increasingly allow for automated modelling (Chowdhary, 2020; Reimers and Gurevych, 2019; Qin et al., 2024; Zhang et al., 2026). Enabling overall assessments of animal welfare is also called for by the growing need for integrated sustainability assessments, where animal welfare to date is only a component besides other social, environmental, and economical components (Bracke et al., 2023). In order to provide a decision support tool that could synthesise the available scientific knowledge about farm animal welfare, a formalised and systematic procedure called semantic modelling of animal welfare has been developed. This procedure can, in principle, incorporate all relevant scientific studies obtained in a broad literature review (Bracke, 2001, 2008; Bracke et al., 2008a).

Semantic modelling is a method to systematically extract and analyse statements (quotes) from the scientific literature that describe how animal welfare is affected under different housing conditions in order to generate an overall welfare score. The term “overall” is used here to denote that all aspects (i.e. all welfare needs of the animals) are considered, as well as all available scientific information (i.e. all relevant and meaningful scientific statements). Compared to a meta-analysis or systematic review, semantic modelling does not require as detailed quantitative reports of, for example, effect sizes or similar numerical and statistical parameters, which often are not systematically reported. Instead, semantic modelling procedures enable a data extraction from the scientific literature using a decomposition of the scientific statements based on an understanding of their meaning in the field of welfare science. Through modelling, written scientific statements are systematically converted into welfare-relevant attributes of housing systems. Different housing systems consist of attributes that can each have an effect on the welfare of animals. Examples of welfare-relevant attributes may include the type of floor, water provision system, and lighting system. Each of those attributes have so-called attribute levels. So, for instance, the water provision system can exist out of drinking nipples, open bowls, or troughs. The attributes receive their weighting factors and so-called attribute level scores based on the science-based ranking of the levels of each attribute in the welfare model (see Table 1 for a glossary). The model takes a description of a housing system as input to produce a welfare score calculated as a weighted average of the attribute level scores and the weighting factors of the attributes (welfare indicator variables) that make up the model. The definition and specification of the welfare-relevant attributes in a semantic model depend on the meaning and systematic decomposition of the scientific statements collected in the database. So, semantic modelling is a method to assess descriptively the overall welfare state as a score on a scale (e.g. 0–1 or, in previous models, 0–10). This is done in a formalised way for a type of (farm, zoo, wild, or pet) animal over a period of time in an assessment domain. The method is based on scientific knowledge with the objective to improve moral and political decision making and sustainable development. In semantic modelling, welfare is typically defined as the quality of life as perceived by the animal itself (Bracke et al., 1999), i.e. as the degree of satisfaction and frustration (i.e. positive and negative affect) of the animal's welfare needs. These needs reflect the state of the animal's cognitive–emotional control systems (which roughly correspond to its behavioural systems or life functions). The overall welfare score (OWSc; see Table 1) estimates, based on available science, the actual welfare state of the animals in a given housing (and management) system in relation to what is theoretically possible. Thus far, semantic models for farm animal welfare assessment have been made for various species and/or specific production directions, such as pregnant sows (SOWEL, Bracke et al., 2002a, b), laying hens (FOWEL, De Mol et al., 2006; Shimmura et al., 2011), dairy cows (COWEL, Ursinus et al., 2009), Atlantic salmon (SWIM, Stien et al., 2013; Pettersen et al., 2014; Folkedal et al., 2016), and rainbow trout and pikeperch (Tschirren et al., 2021). In addition, semantic models have been constructed to assess enrichment materials for pigs (RICHPIG, Bracke, 2008), to assess the risk of tail biting in pigs (PIGTAIL, Bracke et al., 2004a, b), and to assess the severity of handling (esp. sea lice treatments) in Atlantic salmon (Helge Stien et al., 2025). Finally, semantic modelling principles have been used to assess the welfare of calves (Bracke et al., 2008b) and broilers (Bracke et al., 2019) at the system level, the need for wallowing behaviour in pigs (Bracke and Spoolder, 2011), and cross-species comparisons of animal welfare (Bracke, 2006).

In this paper, we present a semantic-modelling framework that further improves the main methodology previously developed in semantic modelling. This work is part of an applied project assessing existing housing systems in Germany (https://www.ktbl.inkalktier.de, last access: 3 July 2026) using semantic modelling on the basis of the SOWEL model. Although the basic principles of semantic modelling were described earlier and were presumed to be clear and bio-logical, we found it was necessary to further clarify and refine the rules to allow new modellers to understand and apply the state-of-art semantic-modelling methodology. This new framework is geared towards a welfare assessment of all major production directions in both existing and novel housing systems. To this end, we developed the ANyWEL (ANimal WELfare assessment of anY farm animal) framework. It can be used to calculate the overall welfare score (OWSc) of any housing system for any production direction once the current scientific knowledge base has been modelled. To this end, artificial intelligence (AI) tools are being developed (see Zhang et al., 2026). As semantic modelling may appear to be a rather abstract activity that is harder to grasp than previously thought, the objective of this paper is to explain the semantic ANyWEL model framework in more detail with the help of a fictitious species called anYmal. In addition, we evaluate the strengths of the ANyWEL model (compared to the original SOWEL model) and its potential limitations. In the companion paper (Part 2, Vonholdt-Wenker et al., 2026), we explain in more detail how scientific statements are to be decomposed formally such that attribute levels can be ranked for welfare based on the scientific information contained in the scientific statements.

To date, semantic welfare models have mainly (but not exclusively) been developed for welfare assessment at the system level rather than, for example, at the farm level (with the SWIM models being a notable exception to this rule). Typically, semantic models consisted of a relational database that stores (i) scientific statements collected from the literature, (ii) descriptions of typical benchmark housing systems covering the assessment domain for which the model was designed using (iii) so-called welfare-relevant attributes (or welfare indicators, such as the attribute of group size), (iv) the ranked attribute levels (e.g. the values of small, medium, and large group size, which specify the welfare-relevant properties of the housing systems), and (v) the welfare needs of animals (e.g. the needs for social contact, movement, safety, thermal comfort) as a basis for the calculation of the OWSc for each housing system. The OWSc is calculated as the weighted average score of the model's attribute level scores (ALevScs, Fig. 1 and Table 1). Model attributes have weighting factors (WFs) which are derived from the scientific statements using a classification involving (vi) so-called weighting categories (WCats, which are classes of welfare measurements) and (vii) weighting-category level scores (WCatLevScs). The basis for modelling is the WCatLevSc. It essentially represents a statistically significant welfare difference between two attribute levels (one of which is better for welfare than the other). This difference generates the ALevScs and the WFs of the attributes that, in the end, make up the welfare model (Fig. 1).

To illustrate the principles of the new ANyWEL model framework, we introduce a fictitious species called anYmal. This is a simple species, kept in only three different housing conditions. The example illustrates the modelling principles with a limited number of welfare needs and model attributes. We use this simplified and imaginary example also to show how the new ANyWEL framework compares to the original SOWEL modelling rules. The illustration is intended not to empirically validate the new model but rather to illustrate how semantic modelling works and to demonstrate the added value of the refined procedures proposed in the ANyWEL framework.

Before we illustrate how to calculate welfare for the imaginary anYmal species, we refer the reader to Table 2, describing the three housing systems (intensive, I; semi-intensive, S; and extensive, E). For these housing systems, we calculate an overall welfare score (OWSc) based on anYmal welfare science. The calculation steps are based on the principles formulated previously for the SOWEL model (Bracke et al., 2002a, b). As illustrated in Fig. 1, the model contains attributes with levels, attribute level scores (ALevScs), and weighting factors (WFs), such that OWScs can be calculated for each housing system. To this end, the model user needs to specify for each attribute which attribute level is applicable to the housing system of interest. Since all attribute levels have been assigned attribute level scores (ALevScs, based on their welfare ranks) and since weighting factors (WFs) have been calculated for attributes in the modelling process, the model allows the calculation of OWScs for the housing systems. The following formulas are used to calculate the OWScs:

Attribute level scores (ALevScs). The ALevScs are calculated for each attribute level on a scale between 0 (worst level) and 1 (best level) proportionally to their rank (Bracke et al., 2002b; De Mol et al., 2006) (see Eq. 1a and Table 5). 1aALevScij=NLi-RLijNLi-1 In the above, ALevScij represents the attribute level score of attribute level j of attribute i. The total number of levels within attribute i is denoted by NLi. The rank of attribute level j of attribute i is denoted as RLij (see Bracke et al., 2002b). The attribute of enrichment material may serve as an example. This attribute has three levels ranked from best to worst: rope, ball, and chain. Here, Eq. (1a) generates the ALevScs of 1, 0.5, and 0, respectively, as shown in Table 5, using Eq. (1b–d): 1bALevSc (Rope)=3-13-1=1, 1cALevSc (Ball)=3-23-1=0.5, 1dALevSc (Chain)=3-33-1=0. The general rule is that intermediate levels (between best, 1, and worst, 0) are given intermediate scores proportional to the number of levels on the scale (0–1). For an attribute with one intermediate level (and three levels in total), the intermediate level gets an ALevSc of 0.5, whereas two intermediate levels get ALevScs of 0.33 and 0.67, respectively, depending on their rank, etc. Note that this equidistance principle between levels is a rule of thumb. The relative distance between any two levels of an attribute containing more than two levels can, in principle, be derived in a way that is similar to the procedure that is used to derive WFs (which reflect the relative distance between the best and worst levels of each attribute). Note, furthermore, that since rope is not actually provided in any of the three existing housing systems (I, S, and E) of anYmals, this best level (rope) of the attribute of enrichment material should be eliminated from the model when the model is designed to only apply to “existing systems” (i.e. existing in the fictitious example). In that case, ball would be the best level and, hence, would have an ALevSc of 1 instead of 0.5. This illustrates how the definition of the assessment domain may affect the model's properties (attribute levels, ALevScs, and WFs of the attribute of enrichment material).

Weighting factors (WFs). The WF denotes the relative importance of an attribute compared to the other attributes in the model. The greater the WF, the more the attribute is contributing to overall welfare (which is calculated as a weighted average of the ALevScs). The WF can be calculated technically as the difference between the sum of the welfare benefits (minus the welfare disadvantages, if there are any) of the best level and the sum of the welfare disadvantages (minus the benefits) of the worst level of the attribute. This is represented by the difference between the sum of the maximum (highest) WCatLevScs of the best level minus the minimum (i.e. lowest) values of the negative WCatLevScs of the worst level of the attribute. In addition to the highest and lowest WCatLevScs per WCat, some additional weight is generated by the number of different types of welfare measures (called types) per WCat. This reflects the idea that different welfare measures add weight, especially across disciplines (rather than within WCats), and that, as such, repeated confirmation using exactly the same welfare measures does not add any weight. Scientific statements may load on one attribute level (compared to another level) with different types of welfare measures within the same WCat, e.g. different types of disease, disease symptoms, or immunological measures all classified in the WCat of illness. For instance, in the example statement 10 (“AnYmals kept in groups of 3 showed a much higher number of life-threatening diseases, such as bacterial and viral respiratory infections due to a low immune response, than in groups of more than 10 anYmals”, Table 3). Bacterial infections and viral respiratory infections are two different types of welfare-relevant measures of the WCat of illness. The inclusion of such different types of welfare measures about the same (best or worst) attribute level should affect the WF of the attribute (and, hence, the overall welfare score, OWSc) because, other things being equal, such additional information (i.e. two types rather than one) should make that attribute a little more important. Therefore, the weightings of the best and worst level can be calculated from the highest and lowest WCatLevScs per WCat, respectively, and from the number of different types (of welfare measures) per WCat assigned to the best and worst attribute levels (Fig. 1). Consequently, the weighting factor (WF) can be calculated as follows (Eq. 2a): 2aWFi=(∑WCatMax(WCatLevScWCat)+∑AD.Type(Max WCatLevScWCat×0.2)Type)ALbest-(∑WCatMin(WCatLevScWCat)+∑AD.Type(MinWCatLevScWCat×0.2)Type)ALworst. Here, i.e. in the ANyWEL framework, WFi represents the absolute weighting factor of attribute i. WFi is the difference (i.e. the middle min (–) sign in the formula) between two terms (between the larger round brackets in Eq. 2a). The first term (a) is the sum (plus sign) of the highest weighting-category level score (WCatLevSc) of each weighting category (WCat) of the best attribute level (ALbest) supplemented with one-fifth (0.2 times) of the highest WCatLevScs of any additional and distinct type of welfare measure for each WCat related to the best attribute level (ALbest). In Formula 2a, the term AD type stands for each additional (A) and distinct (D) type of welfare measure. The second term (b) is the sum (second plus sign) of the lowest (minimum) WCatLevSc of each WCat associated with the worst attribute level (ALworst) and 0.2 times the most negative (min) WCatLevSc of each additional and distinct type associated with ALworst. This differs from the SOWEL model, where, for every distinct type within a WCat, 0.2 was added to the WF regardless of the WCatLevSc associated with that type (of welfare measure).

The new ANyWEL rule is a clear improvement. This can be illustrated by the attribute of enrichment material (for anYmals) with its ALbest rope and ALworst chain. Here, the outcomes of a choice test (statement 14) imply for ALbest rope a WCatLevSc of +3 for the WCat of preference and a WCatLevSc of -1 for the WCat of frustration and avoidance for ALworst chain (Table 5). In statement 14, there were no (i.e. 0) additional types described for ALbest rope. So, there is no added weight derived from any additional and distinct types of welfare measure. Therefore the sum of WCatLevScs for this ALbest is only the +3 for preference. For ALworst chain, there are two types of welfare measure belonging to the WCat of frustration and avoidance as statement 12 (Table 3) reported that anYmals preferred balls over chains based on the time spent playing (i.e. voluntary interaction with the material). Furthermore, mouth and teeth injuries were reported in anYmals when chains were offered (statement 13, Table 3), indicating two more types for the WCat of pain, both of which were scored with a WCatLevSc of -1. Hence, the sum of WCatLevScs for ALworst chain is calculated as (-1Frustration)+(-0.01Frustration×0.2)+(-1Pain)+(-1Pain×0.2) (see Sect. 4 and Vonholdt-Wenker et al., 2026, for details about assigning a ±0.01 WCatLevSc). Now, the WF for the attribute of enrichment material can be calculated as follows (Eq. 2b): 2bWF(Enrichment)=((+3Preference)+(0×0.2))ALBest(Rope)-((-1Frustration(choice test)+(0.2×-0.01Frustration (time spent playing)))+(-1Pain(teeth injury)+(0.2×-1Pain(mouth injury))))ALWorst(Chain)=+3ALBest(Rope)-(-2.2ALWorst(Chain))=5.2. The WF of 5.2 for the attribute of enrichment material means that the best level of that attribute (rope) with an ALevSc of 1 is counted 5.2 times in the calculation of the OWSc.

Additional examples of WF calculations can be found in Supplement S1. Generally, the meaning of a WF is that the higher the WF of an attribute is the more important the attribute is for welfare relative to the other attributes in the model. After the absolute WF for each attribute has been calculated, we sum up all the WFs of all attributes in the model (∑WF). As can be seen from Table S1A1 in the Supplement S1, the attribute of group size has a very high WF. Table 3 shows the imaginary scientific statements on which this is based. Table 4 shows the decomposition of statement 8, reporting often heavy fighting and elevated mortality in groups of three compared to in groups of five or eight. This is linked in Table 4 to three WCats, namely aggression, pain, and fitness, because that is what (we believe) this finding means for welfare (and this is an important fact to take into account). The associated WCatLevScs then generate a relatively higher WF for this attribute (group size) relative to the other attributes (space allowance and enrichment material).

For our anYmals in the assessment domain, the weighting factor sum (∑WF) is 32.92 (see Table S2B1 in the Supplement).

Weighted attribute level scores (WALevScs). The next step in model construction is to calculate for each attribute level its weighted attribute level score (WALevSc). This is done by multiplying the ALevSc (Eq. 1a) of each level of attribute i with the absolute WFi, i.e. the weighting factor of attribute i (Eq. 3a). 3aWALevSc=ALevSc×WFi For example, to calculate the WALevSc of the attribute level of ball of the attribute of enrichment material, the WF of this attribute (i.e. 5.2) is multiplied by the respective ALevSc (i.e. 0.5 for Ball) (see Eq. 3b): 3bWALevSc(EnrichmentBall)=0.5×5.2=2.6.

Overall welfare score (OWSc) per housing system. OWScs can be calculated for each housing system in the model's assessment domain by identifying exactly one attribute level for each attribute in the model (that best describes the housing system), identifying its WALevSc, adding all WALevScs up (∑WALevSc), and dividing them by ∑WFi (Eq. 4a). In this way OWSc is calculated as a weighted average of the ALevScs. The OWSc of a housing system (HS) can range between 0 and 1, where 1 represents the best possible housing condition and 0 represents the worst according to the model. 4aOWScHS=∑WALevScHS∑WF For anYmals, we calculated an OWSc of 0.84 for the intensive housing system, 0.58 for semi-intensive housing, and 0.08 for the extensive system (Eq. 4b–d; S2 in the Supplement). Alternatively, each WALevSc (of all levels of all model attributes) can be divided by ∑WFi separately, even before we calculate OWScs. These could be called normalised WALevScs. Adding up the normalised WALevScs for all of the attributes of a housing system adds up to the OWSc (as e.g. shown in Fig. 2, where the ideal system shows the maximum contribution of the best level of each attribute to overall welfare). 4bOWSc(Intensive)=27.732.9=0.84 4cOWSc(Semi-intensive)=19.032.9=0.58 4dOWSc(Extensive)=2.632.9=0.08 Based on these calculations and the underlying scientific statements about the welfare of anYmals, the intensive housing system is a more welfare-friendly system for anYmals in that it better meets their welfare needs compared to the semi-intensive and extensive systems (Fig. 2). This counter-intuitive result is the direct consequence of the fictitious scientific knowledge postulated about these imaginary anYmals (Table 3). This also highlights that modelling approaches can reveal patterns that are counter-intuitive to experts and may thus help in unravelling novel directions for improving the welfare of farm animals (see e.g. Ursinus et al., 2009; Bracke and Spoolder, 2011).

Overall, the anYmal example demonstrated how the various model components (i.e. scientific statements, needs, attributes, attribute levels, attribute level scores, weighting categories, and weighting-category level scores) and underlying procedures and calculations within the semantic ANyWEL model framework allow the construction of welfare models based on scientific information for different species of farm animals. With these models, users may calculate overall welfare scores for housing systems in the model's assessment domain. During the development of the ANyWEL model framework, the calculation rules were improved. In particular, the rules for weighing the types of welfare measurement and assigning WCatLevScs were adjusted to improve the overall welfare assessments. These refinements will be explained in the next section (Sect. 4).

The objectives of this paper were (i) to explain the general process of semantic modelling and guide the reader through the improved semantic ANyWEL framework with the use of a fictitious animal species called anYmal and (ii) to reflect on the strengths and potential limitations of this model. We have shown how the ANyWEL framework can calculate overall welfare scores (OWScs) for different housing systems in a standardised and formalised way across species using an exemplary anYmal and postulated scientific knowledge to show how the scientific information about the species determines the outcome (rather than what may be expected based on casual generalisation from existing species). Below, we discuss refinements that were implemented in the ANyWEL model compared to previous semantic models (Bracke et al., 2002b) and evaluate the strengths and limitations of our semantic-model framework to indicate why ANyWEL is a major step forward in semantic modelling. In essence, the ANyWEL framework was developed based on the initial concept of the SOWEL model (Bracke et al., 2002b). However, several notable refinements were made by us to enhance the accuracy of the welfare assessment, i.e. the calculation of OWScs.

The ANyWEL framework described in this paper was designed as part of the InKalkTier project (https://www.ktbl.inkalktier.de, last access: 3 July 2026) to provide guidance to new modellers. For the first time, statistically significant welfare findings were presented as the basic unit of overall welfare calculations (by ranking of attribute levels leading to attribute level scores and through the assignment of WCatLevScs). Moreover, the ANyWEL model framework can be used to assess the welfare of any farm animal species. Previous semantic welfare models, such as FOWEL (De Mol et al., 2006) for laying hens, SWIM 1.0 and SWIM 2.0 (Stien et al., 2013; Pettersen et al., 2014) for Atlantic salmon, and COWEL (Ursinus et al., 2009) for dairy cows, had been developed, with modifications, for specific production directions. The ANyWEL approach now allows modellers to assess overall animal welfare in various production directions within the same framework. This is expected to further enhance modelling efficiency and standardisation. Moreover, compared to previous models, we believe ANyWEL is better suited to assess innovative housing systems that lack scientific studies. For this, refined rules were suggested for the incorporation of expert opinions. By including the refined WCatLevScs for weaker statements expressing opinions and tendencies, the ANyWEL model should be able to include more knowledge and calculate animal welfare more accurately. Another advantage of this innovation is the possibility to further study the similarities and differences between welfare assessments based on the published scientific literature and based on “mere” (expert) opinions. This could potentially help bridge the gap between science and society. A major strength of the framework, however, is that it is explicitly and formally based on (the collection and decomposition of) scientific knowledge. A possible weakness associated with this is the existence of knowledge gaps (i.e. gaps in the literature and in what can be retrieved from expert opinion). Such gaps could potentially negatively affect the weighting factors (WFs). As stated above, the WF of an attribute represents how much this attribute contributes to the welfare of an animal (i.e. how important the resource is to the animal and its welfare needs). However, the WF can only be calculated through collecting published or elicited scientific statements, resulting in ranked attribute levels and WCatLevScs (Bracke et al., 2002b, 2008a; Ursinus et al., 2009). Some attributes or interactions between attributes may affect welfare without having been duly considered in scientific studies (e.g. water trough design for dairy cattle seems to be understudied; Pinheiro Machado Filho et al., 2004). The ANyWEL framework can handle these issues via the adjusted WCatLevScs for expert opinions. Expert opinions could help fill these publication gaps and thereby reduce the risk of underestimating such welfare-relevant attributes (provided there is sufficiently reliable expert opinion to justify them). However, remaining knowledge gaps that cannot even be solved with expert opinions remain an issue for science to solve rather than something to solve in semantic modelling as such. Semantic modelling can help to identify knowledge gaps, but it is up to science to close these gaps. Once that has been done, existing models can be updated to accommodate the state of the art. The users of semantic models should be aware that attributes may vary in terms of the number and quality of the scientific (or expert-opinion) statements they are based on. This can be mitigated by the transparency provided by the model, as well as by the fact that limited scientific evidence will generally result in lower WFs. Perhaps we can briefly illustrate this: in our imaginary example we referred to enrichment materials like chains and balls. Within the field of pig welfare science, such enrichment materials are now increasingly regarded as (very) insufficient, perhaps not even worthy of being called enrichment material. For such reasons, it is important that modellers conducting the literature searches have a basic understanding of welfare science generally. New studies showing that, for example, straw and peat are much more effective in activating exploratory behaviour in pigs (e.g. Kauselmann et al., 2021) should, however, fairly “automatically” (i.e. when our semantic-modelling guidelines are adhered to) lead to chains and balls acquiring much lower attribute level scores and, hence, contributing less to overall welfare, thus establishing the view that these materials are hardly enriching for the pigs.

To the superficial reader it may seem that the number of scientific statements positively correlates with the weighting factors. However, it is important to note that the WF represents not the number of statements but their scientific meaning: only the maximum and minimum WCatLevScs within a WCat are used in the calculation of the WFs, and in SOWEL only the distinctly different types of welfare measure generate (a little bit of) extra weight. In ANyWEL, however, the other WCatLevScs are also factored in, but only when it concerns additional and distinct (types of) welfare measures. Hence, a higher number of scientific statements can result in a higher WF only when different measures or larger effects have been reported (Ursinus et al., 2009; Bracke, 2008).

Another point to note is that the output of the semantic model should be cross-checked with current practices on farms using, for example, an overall welfare assessment with animal- and environment-based welfare indicators (e.g. those compiled in the Welfare Quality^® project, Blokhuis et al., 2010). Other ways of validating the model and its attributes are by using sensitivity analysis, a comparison with another set of scientific knowledge (e.g. scientific studies reported in different periods of time, or in even versus uneven years of publication (Bracke et al., 2004a)), on-farm evaluation (Folkedal et al., 2016), and expert opinion (Bracke et al., 2002b, 2007, 2008a). The ANyWEL framework would certainly benefit from further validation, e.g. by a comparison with expert opinions (e.g. on its modelling principles) and application by independent (groups of) researchers. Even without such validation, semantic models can be used tentatively for welfare screening and the design of new welfare-friendly systems (as has, for example, been done with the COWEL model of Ursinus et al., 2009). For some of these applications, such as the design of new systems, assessments at the system level, and integrated sustainability assessments, semantic models may be the only transparent science-based methods available, though external validation remains desirable. For other applications, like on-farm assessments, a comparison (and validation) may be warranted with other instruments like the Welfare Quality^® protocol or qualitative behaviour assessment (QBA, e.g. Cooper and Wemelsfelder, 2020), each of which has advantages and disadvantages in terms of validity, reliability, and feasibility. Note that the relationships with some welfare measures that seem to be able to assess overall welfare like judgement bias tests (JBTs, Neuhauser et al., 2023) and telomere length are multifaceted. Such measures can both be used to validate semantic models (and vice versa) and be classified as other welfare measures and, as such, be used to improve semantic models like other newly available welfare measures or studies would (e.g. JBT may be seen as a special kind of preference test (WCat: preference), while telomere length may be classified in the WCat of fitness). A difficulty here is that, for example, a positive JBT (i.e. showing that anYmals in an enriched environment are more optimistic) generally does not (yet) allow a specification of which properties (attribute levels) of the enriched condition result in the optimism measured with the JBT.

The procedure of semantic modelling itself is highly formalised, comprehensive (taking into account all available knowledge), and designed to take the modeller's point of view out of the evaluation, as has been described in other papers (Bracke et al., 2002b, 2008a). Still, a main criticism in semantic modelling remains the underlying subjectivity (Stien et al., 2013). That problem, however, is inherent in any welfare assessment and even in any welfare experiment, as long as we do not have a valid and direct welfare “thermometer”. For the time being, subjective interpretation of information is necessary to assess animal welfare, and, accordingly, no welfare assessment is value free. Semantic modelling is designed to make the best-possible assessment based on all available and reliable information, i.e. on scientific statements collected in a database. In semantic modelling, the modellers decide which WCat and which WCatLevSc they assign to which attribute level when decomposing a scientific statement. In applying a semantic model, the users must also decide which welfare-relevant attribute levels apply in the housing systems they want to assess. We cannot rule out that modellers and users can make mistakes. Semantic modelling, however, also allows for making all decisions transparent, such that errors can, in principle, be identified, discussed, and corrected. In case of ambivalence, it is possible to run alternative model calculations (e.g. with or without WFs) and settle disputes. For example, sensitivity analysis of existing models, which usually have more than 30 attributes, showed that OWScs calculated with and without WFs (i.e. with all WFs set at 1) were highly similar in terms of output (i.e. OWScs of benchmark systems). This suggests that disputes about the assignment of WCatLevSc may have a relatively minor impact compared to, for example, the number of attributes identified in the model (Bracke et al., 2002a, b; Ursinus et al., 2009). In semantic modelling, the impact of subjective decisions also decreases as the quantity and quality of the scientific statements collected from peer-reviewed publications and other reliable sources increase (Stien et al., 2013). To further formalise and standardise the semantic-modelling procedure of selecting and decomposing scientific statements and objectively weighting the welfare effect, a guideline to enhance inter-modeller reliability has been developed (Vonholdt-Wenker et al., 2026).

In contrast to existing semantic models, which included housing, as well as management, factors in the assessment domain (Bracke et al., 2002b; De Mol et al., 2006; Ursinus et al., 2009; Stien et al., 2013), the ANyWEL model framework was designed to apply to different species, housing systems, and assessment domains, resulting in a specific welfare output which we labelled “welfare potential”. This was done because, for the InKalkTier project, in the context of which ANyWEL was developed, management factors were not to be included in the welfare assessment (i.e. InKalkTier assumed best management practices) – hence the term welfare potential rather than actual or realised welfare. However, we think that the ANyWEL framework can also be applied to construct models that include suboptimal management practices.

The ANyWEL framework, furthermore, allows the development of models for different life stages of the animals, e.g. for pigs in the farrowing, weaned, growing, and fattening stages. In these cases, for example, the available amount of space per pig changes over time (Chidgey, 2023). When the aim is to make the best-possible welfare assessment throughout their entire lifetime, perhaps even including the transport and slaughter phases, surely future work needs to be done to integrate the output of the different component models. In addition to that, further work needs to clarify and formalise the process of combining attribute levels (e.g. combination of perch height and perch shape for laying hens).

Beyond farm animals and welfare, the ANyWEL model framework may be extended to cover other species, e.g. zoo animals, wild animals, laboratory, and companion animals, and perhaps also other (social) sustainability aspects (which seem mostly centred around human welfare). Thus, the suggestion we make here is that the semantic-modelling approach could also go beyond animal welfare to assess the sustainability, such as in terms of environmental and economic impacts, of livestock farming (Bracke et al., 2023). This may makes sense once we realise that ANyWEL is to apply to any animal and that what matters for sustainability is to allow humans and animals to thrive and experience positive welfare (Rault et al., 2025). The ANyWEL framework remains to be implemented in actual welfare models, but welfare statements about various species of livestock have been extracted using the ANyWEL framework (Benthin et al., 2023). These have been successfully used to develop an AI tool for automated statement extraction (Zhang et al., 2026). This is in line with the idea that the principles underlying semantic modelling are bio-logical, i.e. based on (an understanding of) biology and logic (see the explanation of using if–then rules in Vonholdt-Wenker et al., 2026).

ANyWEL is a novel framework for assessing the welfare state of different species and production stages of farm animals in relation to housing (and management) based on scientific statements. The framework prescribes how to integrate scientific findings to generate welfare scores to compare and benchmark different systems. The ANyWEL framework can be used to make semantic models to calculate overall welfare scores for existing and new housing systems. This can improve insight into the relative contribution of different aspects (attributes) to overall welfare and the design of new welfare-friendly systems. The greatest advantage of the ANyWEL model framework is its ability to standardise welfare assessment across species. Compared to the original semantic model (SOWEL) on which ANyWEL is based, the accuracy of the modelling framework was improved. To this end, we adjusted the calculation rules (e.g. for types of welfare measures and the assignment of weighting-category level scores). The ANyWEL framework was illustrated using a fictitious animal species called anYmal kept in three different housing systems (intensive, semi-intensive, and extensive). We showed how expected and subjective welfare assessment may be modified once scientific knowledge is considered explicitly following the ANyWEL procedures.