In accordance with Firth’s and Harris’ intuitions, the algorithm works by considering the contexts of word-types (by looking at what word-types surround them) and saving that data in a proper data-structure – meaning that word-types will be represented as the other words that they co-occur with and the counts of these co-occurrences. First we must get the base information we need from the texts. This is done by counting for every word-type which other words co-occur with them within some window of size n. Here n signifies the number of words that we look at on both sides of all the occurrences of the word-type investigated. This means that if we for example choose a window of size 4, and consider the following sentence:

We can derive a representation of this sentence in terms of relative closeness of word-types. There are 12 different word-types in the text. For each word-type we can ask how often each word-type occurs (given a particular windowsize). In table 1 below you can see what values this would deliver (the first column shows how many occurrences of the word-type (WT) occur in total):

This is a vector-representation of the contextual information in the text. Each of the rows is a vector (i.e. there is a vector for each word-type) in a 12-dimensional space. Every individual number in the table represents how often the column word co-occurs with the row word. Important to note is that the windowsize directly impacts the nature of the results of the vector-representation of the text – in our experiment in section 5, we use a windowsize of 10. There is a strong case to be made that taking different values for the windowsize can in many ways be used to extract not better or lesser results, but properly different results from the text.[2]

However, these raw counts are, on their own, not particularly informative. They need to be transformed so that we can extract values that indicate how strongly two word-types are connected that is not directly influenced by the total occurrences of certain word-types. We see in the above table for example that ‘movement’, ‘is’, and ‘difficulties’ all have higher total co-occurrence counts than the other words – a direct consequence of their own higher frequency. However, it is not so that by occurring more often, a word-type is necessarily more connected to more types of words in interesting or salient ways. This means we need to transform these raw counts-scores into a score that is unaffected by total frequency of word-types. For this we have used and will now introduce the measure pointwise mutual information.

Each of the pairs of word-types needs to be scored – or, their connectedness needs to be measured – based on the data from Table 1. The way to do this is by extracting the data about the number of times they have co-occurred, together with how many instances of the word-type there were. The method we will be using to do this is Pointwise Mutual Information (PMI).[3] PMI measures the chance of finding a word-type y within a window around another word-type x (within a chosen windowsize) in the investigated text, divided by the chance of finding y in the whole text. The idea is that related words should be found disproportionately in each other’s windows[4]. The overall chance of finding y indicates a regular proportion. We take the log of this value in order to make the results better suited for vector calculus later. We will write small x for the set of instances of x itself and capital X for the windows around instances of x. Notice that the different windows X might overlap. In this case the same word-token will be counted twice or more. Hence X should be seen as the disjoint or indexed union of the windows. The formula for PMI reads:

P M I x , y = l o g 2 P y ∈ X P ( y )

This provides us with the average of the chance of finding a token of y in a window near a token of x, divided by the overall chance of finding y in the text. The fraction can only be a positive value (both denominator and numerator are positive), but a log of a value between 0 and 1 gives a negative value. If the two word-types are fully independent from each other, we’d see that there is no difference in the chance of finding x and y in window-proximity to each other in our actual text, versus the hypothetical situation where they have no actual relation – i.e. the fraction will then give as a result 1. Log(1) = 0, so PMI scores above 0 should be read as indicating a positive correlation between the two word-types in the text, whereas scores below 0 indicate negative correlation, and 0 itself indicates independence.

Computationally the chance P(y) is simply the instances of y divided by the total word count. The chance P y ∈ X is found by counting each instance of y within the window of an instance of x, where the same y may be counted more than once when it appears in more than one window, divided by the size of the disjoint union X of all the windows, which will always be the window size multiplied by the amount of times x appears in the text.

p y ∈ X P y = | y f o u n d i n w i n d o w s a r o u n d x | | x | * w i n d o w s i z e | y | T o t a l w o r d s

So for example, when we consider table 1 above, we can see that the score for PMI(is, with) with a window size of four is given by the fact that we find ‘with’ around ‘is’ twice, that ‘is’ occurs twice itself, that ‘with’ occurs once and that the entire text consists of 15 words, giving us:

l o g 2 2 2 * 4 1 15 = 0.25 0.0667 = 3.75

When the corpus is only a single sentence, this method does not yet pick up on any significant semantic properties – but as we will see later on, when the number of sentences and words, grow, this problem diminishes.

From this vector-representation of a word with scores based on PMI, there are (at least) two routes available for the extraction of different meaning-representations. We argue that one discovers more salient aspects of meaning, whereas the other discovers the more subtle/less salient aspects of meaning.

Many of the scores inside the vector representation are low. The connectedness of the terms is there, but not particularly salient. Specifically within the approach that is called the extraction of collocations, this intuition is put to good use. Collocates to a specific word are those words that are particularly connected to the word that is being investigated. Instead of looking at 100.000 word-types and their connectedness to a word of interest, in these cases, one restricts the list to those word-types that score at least a certain score on the PMI measure. This often leaves investigators with a handful, up to a ~100 word-types, which together provide a good overview of the interesting semantic properties of the word-type under investigation.[5] In our own algorithm, we have taken the threshold to be a PMI-score of 4, so as to mimic these applications, which often aim to consider around a few dozen of word-types for an investigated word-type.[6]

We argue that what investigators making use of lists of collocates are extracting are those connections between word-types that are (within the investigated corpus) particularly salient, and by extension, more public and explicit. This works well for their purposes. The list of high scoring words will mostly include words that, to a reader of, or author in, the corpus, will likely ‘make immediate sense’ when provided. Of course, the nature of the connection won’t be provided, but that they are connected will be clear on most occasions to most people who read the corpus – and looking at this limited set of words can provide an investigator with a lot of useful information on the corpus.

Contrast this with the vector representation. Here the meaning of a word-type is not given by a list of x particularly saliently connected words, but by the PMI score with every single other word-type contained in the corpus. In addition, instead of taking an implicit ‘one-hot’ representation as in the collocation extraction (another word-type either is, or is not, connected enough) the vector representation typically takes the exact outcome of the PMI calculation as its score. This means that even minute differences will play a role in the differentiation between different word-types. What we are picking up on here, we argue, are the more hidden, less explicit, connections a word-type makes within the corpus. It investigates how the particular ways of using a term influence further usage and application of a term, despite it being particularly hard for a speaker or reader to explicate that such a causally productive connection exists.

We have suggested that PMI scores track relevant semantic properties of a word. In addition, that high scores track more salient features, and low scores more implicit and subtle features. What argument or evidence is there for this? Perhaps the use of a word is part of its meaning but surely what other words it appears near in written text is at best an approximation of how the word is used.

Psychological research shows that PMI is positively correlated with judgment of similarity, that is to say if words have higher PMI scores in sample texts then test-subjects are more likely to judge them similar [McDonald 2000, 35–67]. In addition, texts with high PMI scores for given word pairs have been shown to make people see the words as more related in meaning [McDonald and Ramscar 2001]. In addition, as McDonald points out, such scores can be correlated with other psychological variables such as effect in lexical priming [Lowe and McDonald 2000], analogous reasoning [Ramscar and Yarlett 2000] and synonym selection [Landauer and Dumais 1997].

This shows that whatever properties are tracked by PMI are not mere artifacts of the texts studied but have at least some correlation with psychological indicators of word meaning. What does not follow from this work is that lower PMI scores indicate less salient aspects of the use of a word. That would, by its very nature, be hard to extract from straightforward questions to test subjects, though it might be studied in other ways. In section 5 we will include results from an explorative case study that show that in at least some cases, looking at just the high scoring words yields different results from looking at all word associations.

The above methods describe how we propose to be able to extract an approximation of less or more salient aspects of meaning. In the first case, the vector is the representation of the less salient aspects of the concept of a word-type in a text/corpus, in the latter case, the list of connected terms is the representation of the more salient aspects of the concept of a word-type in a text/corpus. In the case of the extracted list of collocates, research sometimes focusses on the qualitative investigation of this list of words itself (indeed, the ability to do so is one of the attractive features of the entire approach).[7] The vector approach generally does not allow for this and is often coupled together with a measure of the similarity between vectors. This allows the researcher, for example, to check the stability of terms over time, find synonyms, and discover dissimilarity between terms which are expected to hang together.[8] We introduce a commonly applied similarity measure for vector-representations and argue that there are useful options for measuring the similarity of lists of collocates. These measures will be used in the case study to extract useful information about our corpus for both sorts of meaning-representations.

In the case of the vector representation, what we are looking for is actually a measure for closeness of coordinates in a multi-dimensional space. A number of different options is available, but one of the more often used measures is cosine similarity [Han 2012, sec. 2.4.7]. Intuitively, how similar two vectors are is measured by looking at how much they coincide for each of their coordinates. A simple measure for this is provided in linear algebra by the inner product, or, dot-product. This measure sums the products of each pair of coordinates that lie in the same dimension, or:

v ⃗ ∙ w ⃗ = ∑ i = 1 N v 1 * w 1 + v 2 * w 2 + . . + v N * w N

Partly, this measure does what we want; the more (high) values shared between two vectors in similar dimensions, the higher the similarity of the two vectors will be. However, what is counter-intuitive is that longer vectors (vectors with higher values in many dimensions) will generally be scored higher than shorter vectors will be. However, it is not the case that just because a word-type has more strong connections with other word-types (higher PMI-scores) it should be deemed to be expected to be more similar to other word-types in general. This problem can be amended simply by normalizing the lengths of the two vectors; in this way, we only measure the similarity in distribution of values across the different dimensions of the two vectors, while negating the influence of their absolute lengths. We get for our similarity measure:

v ⃗ ∙ w ⃗ v ⃗ * | w ⃗ |

Linear algebra shows that this formula captures the same thing as taking the cosine of the angle between the two vectors v and w, hence the measure’s name cosine similarity. Higher values are attained when the distribution of values between two vectors amongst all the possible dimensions is similar – or, when the angle between the two vectors is close to zero (in addition, for orthogonal vectors, we get a similarity of 0, and for opposite vectors, a value of -1). This allows our measure to provide values between -1 and 1.

In the case of the lists of terms, how to model similarity is less researched because lists of terms do not easily translate into a question about vectors in space. Intuitively, however, what we want to measure is the amount of overlap two terms exhibit in their lists of collocates. The greater the overlap, the larger the amount of saliently connected terms two words share. What is dangerous however is that we do not want that word-types that have a larger number of strongly connected terms be more similar in general to other words (which would happen if we would take ‘a large overlap’ to be signified by a large number of overlapping terms). This needs to be normalized. One of the measures used for these purposes is the Jaccard index, which scores the overlap between the neighborhood of two nodes in a network in the following way:

| A ∩ B | | A ∪ B |

The intersection provides the overlap and the union provides the sense of the ‘total’ size of the neighborhoods; sharing a large amount of terms provides a higher score the smaller the total sample size of potentially shared terms becomes.

This measure is straightforward in the special case that the neighborhoods have the same cardinality. The sets {E, F, G, H} and {G, H, I, J} will have a Jaccard index of 1/3; the intersection contains two elements (G and H), whereas the union contains six elements (E, F, G, H, I, J). In the case of the same cardinality, we also have the nice property that complete overlap will provide a score of 1. This breaks down in the case of dissimilar cardinalities. Take the Jaccard Index over {E, F} and {E, F, G, H}. This will turn out to be a half (two shared elements and four total elements). This means that the scores will be influenced by the lengths of the lists of collocates. However, this on its own is not a bad thing, as the length of the list tells us something about how much salient connections there are for the word within the corpus. If two terms significantly differ in this respect, it is reasonable to take this along in the calculations. What we do want to avoid (and what the Jaccard Index avoids) is that words with larger or smaller lists of terms get scored more or less highly on their similarity scores in general. This is not the case, since the length of the list only comes into play relatively to the length of the list of another word-type with which the similarity is measured. Word-types with the same, or a similar, neighborhood size will have a higher potential score, but this does not translate into a preference for either larger or smaller neighborhoods to influence the scoring on its own. In addition, the Jaccard index is commutative. It will always tell us that a first word is as similar to a second word as the second to the first. This maintains the intuitive symmetry of the similarity relation.

As a brief intermezzo we will discuss why our mode of interpretation, which focuses on the salience of individual scores in the PMI-based semantic vector, would be much harder for word embedding based methods like word2vec [Mikolov et al. 2013] and BERT [Devilin 2018]. Such methods have been used in many semantic and linguistic contexts, including recently the tracing of semantic change in historical corpora [Hamilton 2016], [Wevers and Koolen 2020]. In fact, for many tasks like tracking synonymy and homonymy word-embeddings are superior to the sparse vector approaches that have been interpreted above.

Word embeddings can make use of different architectures, but let us look at the continuous bag-of-words model. Making use of a neural net, an algorithm attempts to invent word vectors that are maximally useful in the task of predicting, on the basis of a collection of words surrounding a word’s position (the bag-of-words), what word will occur in the middle of these words. By altering the values (weights) of a word’s entries into its vector so as to be maximally effective at this task, a vector representation of the word is learned. As the algorithm learns, it doesn’t find the eventual weights, or scores, of words along a number of axes, it defines its own, maximally informative, axes. That is to say, in using a method like word2vec, not only are the scores that should define a word-vector extracted, but also the best conceptual frame of representing the word-vectors is constructed. However, due to the difficulty in interpreting what a particular axis stands for, it also becomes difficult to interpret individual scores along this axis as is done above (entries being interpreted for example, as more or less salient connection between word-types).

Consider the vector describing word type ‘The’ in table 1. This vector has 12 values, each representing the number of times the word ‘the’ co-occurred with the 12 other word types. Each of the vectors can be retraced to interpretable quantities through its elements which are interpretable quantities. By contrast a comparable word2vec vector would contain entries that represent variables invented by the algorithm. This does not allow for the word-type-pair-based analysis we propose.

To put it a bit more technically, word embeddings use various techniques to represent semantic properties in lowdimensional spaces. This reduction in the amount of dimensions has computational advantages, but the downside from an interpretative point of view is that it is less clear what the elements of vectors represent (in any case, there is no reasons to expect them to represent word-type-pair relations). Our methods also allow us to define a vector-space and norms on it (see the section on similarity measures) which has far more dimensions, but the directions, vectors, and individual entries can be much more readily interpreted because this space was invented by and for people.

It might be possible, instead of trying to interpret the individual steps of the algorithm produced by a teaching algorithm, to instead interpret what word embedding methods are good at, or what using certain training data teaches them[9], but that is beyond the scope of this paper.

Having discussed two ways to model words, vector-representations and collocate-representation, which respectively model the less and the more salient aspects of its meaning, we now move to the concept of the stability of a word-type within a corpus of individual texts, which is the final ingredient required for our proposed application of combining differently measuring the more and less salient aspects of meaning.

For each of the works in a given corpus we can generate both vector and collocation representations of a set of different word-types that are relevant to the corpus. Using the similarity measures we have discussed above, we are able to compare each of these models to one another. In particular, we compare the models of the same word in different works. For instance, the usage of the term ‘cause’ in one work of the corpus can then be compared on the level of similarity (for both facets) with the usage of ‘cause’ in another work. For our two methods we then get a relation of similarity for each word-type between every work in the corpus. These results can be used for building up a network representation of the corpus. Consider for example Figure 1 below, which provides the resulting graph, based on the network of all similarity scores between all works in the corpus of early-modern English physics/natural philosophy textbooks that will be further detailed in section 5, for the vector-representation based method, indexed on the word ‘body’:

The nodes stand for the different works in the corpus, while the weighted edges that connect them stand for the similarity between the two works in the generated models of the investigated word-type. Of course, one can analyze such a (set of) networks (one for each word-type) in many different ways to allow for interesting corpus analysis.[10] We will only make use of a particular route of investigation – we will look at the average weight of the edges for each of these different networks (i.e. word-types) for both of the models available. On its own, this average of the edge weight intuitively gives a number that indicates how similar the word is used across the corpus. A high score indicates that many of the works have particularly similar models for their particular meaning of the investigated word-type. A low score by contrast tells us that the word is not used particularly similar across the corpus, and that the meaning of the term is particularly unstable. We’ll label these scores ‘stability scores’. Higher values signal that a term under investigation is used more similarly across the entire corpus, i.e., the term’s usage is more stable in that corpus than one that scores lower.

Our important claim regarding these numbers comes from the divergence between stability scores obtained by the same word-type, depending on whether we are considering their more or less salient aspects of meaning. Indeed, whereas one might expect to find two kinds of terms – terms that have high stability scores for both the more and less salient aspects of their meaning, and that score low on both – we find divergence between these scores. It is these divergencies of stability scores when one differentiates between more and less salient aspects of meaning that must be interpreted and that will be provided as our case-study. To investigate such scores we call salience differentiated stability analysis.

In the upper right quadrant we find 4 terms (excluding motion) – ‘earth’, ‘water’, ‘sun’ and ‘object’. The commonality between the first three terms is that they all refer to identifiable, concrete, phenomenally accessible, things. This is of course not to say there were no interesting discussions about these terms (a lot of astronomical work was done on the relation between the earth and the sun, and on the nature of these two planets) but the terms are still relatively straightforward qua meaning. For the sun in particular a simple ostensive definition is available, somewhat similarly for water (water being this ostensively available stuff, and other similar stuff). Indeed, the three terms are all very neatly counted as easily identifiable, concrete, objects of investigation within the discipline of natural philosophy (fire and electricity also inch close to this quadrant). ‘Object’ is the only abstract term of the four and also scores somewhat closer to the façade quadrant than the rest. But, the usage of a term like ‘object’ is simple – and there can be little truly different applications of it. Because even when speaking of mental, biological or physical objects, in all cases nothing changes about the ways in which they are objects.

There are three terms in our case-study that fall (or almost fall) in the upper left quadrant that we have proposed to understand as being comprised of terms that are extremely central to the discourse, are central to much controversy, but enforce a certain unity into the discourse via implicit agreement on the more general connections that such terms have. ‘Motion’ and ‘matter’ are edge cases, so let’s first look at the clearest case, ‘body’.

‘Body’ is extremely central to early modern natural philosophy. The early mechanicist philosophies in the 16th century (like Descartes’ philosophy) let themselves be summarized in the striking sentence “bodies in motion,” or, “matter in motion” [Nadler 1993, 3], [Roux 2017, 27–28]. Although many schools (already existing in the form of scholastic schools, or following schools, like experimentalists and Newtonians) will reject mechanicism’s claim that the investigation of “bodies in motion” exhausts the activities of the natural philosopher, they however will all have to concede that the mechanicist has set the program: to be a natural philosopher is to (at least) have answers to questions pertaining to bodies (whatever they might exactly be) and their movements (whatever that might end up meaning exactly).

Once the program has been set – the natural philosopher is to say something about bodies – it is natural that ‘body’ itself becomes a contested term. The philosopher who can provide and defend a conception of ‘body’ such that it easily fits with his/her more general outlook is a philosopher who has successfully scaled the walls of natural philosophy. At the same time, these moves need to be made out in the open – since the eye of the reader is fixed on these terms. Façade-like behavior is not to be expected – all of the disagreement is out in the open, not hidden, nor is there any manifest agreement that could help hide subterranean disagreement.

A few examples will show the extent of the mutability of ‘body’ within early modern philosophy, where these mutations are central to the arguments provided. Descartes’ mechanicism comes together with a very explicit statement about the nature of what bodies are – they are primarily to be characterized by their extension. This is coupled together with a statement about the other ‘type of thing’ there is in the world, namely, thought. Bodies being extension means that all the other properties of bodies can be explained reductively by reference to their extensional properties (and their motions) – except any properties (or movements) that are to be explained via thought. If body is anything, it is to be extension, and if body is to not be anything in particular, it is not to be thought. Amazingly, in opposition to this, we find later natural philosophers to remain true to Descartes style of reasoning about body (1. Body is central to their activity as natural philosophers and 2. It is an investigation into body’s essential attributes and motions that should occupy them in particular) while in the meantime moving away wholly from his conception of body. Indeed, to make room for their particular systems of natural philosophy, we find thinkers like Moore, Cavendish and Conway, more or less destroy all properties ascribed to bodies by Descartes. For Conway, both the impenetrability (one of the ways the space filling ‘extension’ was often fleshed out) of bodies is opposed and it is claimed that bodies are both spiritual and material-like [Lopston 1982, 15]. Cavendish phrases the issue somewhat differently, but for her too, bodies are coupled together with mental properties [Shaheen 2019, 3553–3554], as is the case for Moore, who arguably provides less reworking since he mostly takes the extended conception of motion to be gappy in accounting for phenomena [Roux 2017, 27]. What allows for these very deep metaphysical discussions is that, in the end, bodies can be somewhat easily identified in everyday activity. And whatever the exact nature of bodies turns out to be, it should in the end still be mapped on a set of paradigmatic instances of bodies (even the radical turnaround one finds in a thinker like Conway, where one would expect the concept to be so definitively deformed to no longer map unto the same objects as previous concepts of body, is still applied to recognizable cases, like the different unmixed fluids and how they hang together in a single bottle).[14]

So, what about ‘motion’ and ‘matter’? Generally, ‘matter’ was used more restrictively than ‘body’ – and was less easily identified with everyday objects in the way that bodies could be. One would expect that matter was less stable in application because matter was not so strictly tied together to phenomenally accessible paradigmatic examples (bodies might be best imagined as spheres, houses, coherent swathes of fluid, cannon balls whilst matter was often understood to encompass more restrictively the corpuscles that make up bodies). This is due to the Aristotelian roots of the concept – matter plays a role in the hylomorphic theories where things are essentially made up out of the combination of matter and form [Manning 2012].

Motion get reinterpreted radically within early-modern philosophy, motion as change makes place for motion as mere local-motion. However, this move is not as extensively challenged in the development of early modern natural philosophy as was body. While body functioned as the pivotal point where different conceptualizations could be made to do all of the metaphysical lifting, motion did not take up such a place. Motion was on the agenda but was more easily understood and less prone to the radical reinterpretations (after the initial reconfiguration) than occurred for body.

In the lower right quadrant we would expect to find terms that exhibit façade like behavior. These are words similar in salient ways but dissimilar in more subtle ways. Superficial similarity may mask slippery differences in usage. Three terms seem to apply – ‘fire’, ‘electricity’, and ‘method’. We will not discuss ‘fire’, because, as we can see in the plot the distance between ‘fire’ and the ‘x=x’ line is very small. That is to say, the difference between the more and less salient scores is very small (98/104).

Similarly, ‘electricity’ will not be extensively considered for two reasons. Firstly, it’s quite close on the line toward the upper right quadrant. In addition, the technical nature of the term precludes a useful analysis without more extensive background knowledge of the term. A few short remarks that should be read in that light are that: ‘electricity’ has properties in common with ‘earth’, ‘sun’ and ‘water’ in the sense that it is a term that designates a somewhat properly delineated group of phenomena. In particular, static electrical effects (which resulted in the attraction of other objects) and magnetic materials (which show somewhat similar behavior) were termed electrical. Although practitioners disagreed on how to explain these phenomena, for some time, this set of phenomena was clearly delineated. However, further study revealed that some previously ‘electrical’ phenomena should be delineated from ‘the electrical per se’, in particular excluding magnetic phenomena from the term's extension [Gregory 2007, 35–42]. In addition, there is the novelty of the topic in natural philosophy as a systematic topic of interest – whereas it was of fringe importance in foregoing natural philosophy, it gets conceptualized more systematically within early modern natural philosophy – the novelty should induce us to expect less stability in the application of the term in its non-salient features, whereas identifiable phenomena should make us expect more stable non-salient applications. What is less easily explained is the cause for the agreement about salient facets of electricity’s meaning – especially since the debate about electricity is often characterized as a number of schools disagreeing fundamentally about the nature of electricity (Ibid, p.35-42). Note that most of the systematic early modern explanation started to come up in the second half of the 18th century, making it only a small part of the timeline of our corpus. We leave this discussion as is.

Central in the lower right quadrant is ‘method’. Many early modern thinkers agreed on the centrality of method in science and presumably its salient stability can be explained by this as well as by a shared understanding of the overall concept. We might be able to explain its subtle instability by the fact that method as a word generally signifies that some technicalities are to come, but which technicalities differs depending on which method in particular will be discussed.

If we are correct, method’s presence in this quadrant can be explained without calling it a façade. Electricity might be a façade, but to uncover this would require more in depth analysis of the development of the concept of electricity in this period. This means that our case-study did not bear out our expectations – no clear cases of facades were found in the lower right quadrant.

The most densely populated quadrant is that of the crisis terms (5 terms, excluding ‘matter’) – ‘part’, ‘cause’, ‘specie’, ‘form’, ‘experiment’. These are words which are understood differently and used differently. These terms are thus such that we should expect both that the terms were controversial and discussed explicitly in natural philosophy and that there were little methods available to tie these terms together via, for example, implicit agreement on the (paradigmatic) extension of these terms (as in the case of body).

All of these terms agree with this characterization. The four terms ‘part’, ‘cause’, ‘specie’ and ‘form’ are all derived from scholastic philosophy and heavily debated in early modern natural philosophy. Species, (substantial) forms and causes were all important aspects of the scholastic/Aristotelian framework, and were all reworked, or even outright rejected, in subsequent schools of natural philosophy (like Cartesianism and Newtonianism) [Blair 2006, 366]. Cause is already reworked by Descartes, who also rejects (although not fully) substantial forms [Flage 1997, 845]. We see species being rejected by mechanicists as well, as well as non-mechanicists like Conway [Conway 1996, 30–31]. Cause is outright rejected as being the proper object of investigation for natural philosophy by later Newtonians like van Musschenbroek [Sangiacomo 2018, 51]. Forms are sometimes reintroduced and reworked and all the while remain in play in the strong scholastic school that remains in operation for most of the early-modern period, particularly in the university context [Sangiacomo et al. 2022b]. What differentiates these terms from body are two things: i) body has easily accessible paradigmatic instances of the concept’s extension which are not so readily available for, for example, part and form (and arguably, at least some of the fourfold of Aristotelian causes) and ii) whereas body was extensively discussed in the light of its given central position to the discipline, these terms were discussed because their position within the discipline were under dispute. Experiment has a similar structure, except there the discussion will not have been most explicitly between scholastic and new philosophies, but between rationalist/Hobbesian conceptions of natural philosophy and experimental/Boylean [Shapin and Schaffer 1985/2018]. Again, not only is experiment a more technical term that allows for less easy identification of its extension (what even counts as an experiment, as opposed to observation, or an uninformative ‘account’) [Shapin and Schaffer 1985/2018]. In these debates, it is also still up for grabs whether it has any place in natural philosophy. In this sense we propose the terms in quadrant to be crisis terms – they are central to the discipline, because in many ways, the form of the discipline is transformed by drastically questioning the contents and validity of these terms.