CEFISES Seminar: Eran Tal, “Measurement, Prediction, and Fact”

[music] Um okay so uh welcome to everyone for the last session of the meisto seminar. Today it is my pleasure to introduce an associate professor at me university. Iran has shaped much of the contemporary philosophy of measurement developing a distinctive model based epistemology and more recently a pragmatist approach to the factuality of measurement results. Today he will offer us a new pragmatist solution to uh the rebal of factuality. Uh so thank you very much Iran for being here and uh the stage is yours. >> Thank you very much Danielle and thank you all for uh for joining today. I hope you can hear me. Okay. >> Um I uh I'm uh I apologize for not being able to uh be with you in person today. Um but I was very pleased to receive the invitation to speak with your group. Um the the title of my talk um is measurement prediction and fact and it's a work in progress that I'm currently um preparing for publication. So any and all comments are extremely welcome. Um the to to start the um there there are two u claims I want you to consider two types of claims um about the values of quantities. Um the top claim um is an expression of a categorical expression of a a claim a factual claim about the temperature. The temperature of object O is T uh with some margin of uncertainty U. Um the bottom claim uh is different. If conditions C1 uh to CN held, the temperature of object O would be T with some margin of uncertainty U. That's a hypothetical conditional claim about what the temperature of object O would be if the conditions held. Usually we think that measurement outcomes, measurement results, the knowledge we get from measuring belongs to the first kind or should be expressed by claims of the first sort. Categorical claims, claims about what the temperature is. After all, thermometers are supposed to tell us what the temperature is, not what the temperature would be under some conditions. Um, and that's how we widely treat measurement outcomes. uh and this this seems to be epistemically important important especially for the ability of measurement outcomes to serve as scient scientific evidence. The evidential power of measurement is supposedly predicated or conditioned on their ability to provide us knowledge of what quantity values are or more generally measured values are >> not just what they would be under some conditions or under some assumptions. Uh but there's a tension between our intuitions about the knowledge measurement provides us and the practice of generating measurement outcomes. The practice of measuring >> measurement outcomes we know are inferred from theoretical and statistical models of measurement processes. And I'll talk more about what that looks like later in the talk. But what [clears throat] I'll try to show you today is that um these models um of measurement processes cannot be independently tested. They cannot be tested independently of some other model of some other process. Um in other words, um measurement outcomes turn out to be the consequences or predictions of um of theoretical and statistical models that are only compared to each other rather than compared to something independent of models. Um whether it be some um independent observation or signal or cause. um at least uh with the quantitative exactness that we we expect for measurement we simply do not get this sort of independent test. So it seems that if we simply uh take the uh if we simply look at the practice of measuring measurement outcomes shouldn't be expressed as as categorical claims about the temperature what the temperature is but rather as hypothetical knowledge claims about what the temperature would be if some model was adequate. And that's a problem or at least it it's a tension between our intuitions about what the measure the knowledge measurement provides and the and the practice of generating inferring measurement outcomes um as we know it. I'm calling this tension or problem the riddle of factuality. Um and I'll uh spend today going through uh a few different ways of solving it. Um there are in in the in the work I'm I'm writing there are four accounts um empiricist, information theoretic, causal and pragmatist. Today uh for u for lack of more time I'm going to skip the information theoretic accounts and focus on first empiricist and causal accounts. I'll show that um persistent causal accounts um as they are uh formulated currently are inadequate solutions to this riddle. They do not provide us uh with an explanation for why we why we should think measurement provides factual knowledge. Um and I'll then then turn to um a pragmatist solution that I think solves the riddle. So that's the plan uh for today. >> Okay. Um so let's start uh by talking about measurement and empiricism. And um a lot of it will be rather broad uh strokes. Um so um I'm I'm sure there are subtleties that uh will come out uh during uh Q&A and I'm very happy to delve deeper into the uh into the into these subtleties uh during the discussion. Um but in broad strokes when we talk about uh measurement all measurement procedures involve some kind of observation uh at least at some point. Um, if you're an empiricist, you're committed to more than just saying that measurement involves observation. Um, if you're an empiricist about measurement, you you would say that measurement is a a type of observation, a rigorous type of observation. Um, think about what you do when you um when you use a clock. you you count the ticks of a pendulum clock, for example, or you observe the um the height of a mercury uh column or the height of water in a measuring cup to uh to measure volume. Um if you're an empiricist about measurement, you take uh measurement to simply be a more disciplined uh rule-based type of observation. the empirical content of a measurement outcome is determined by the observations you make. Um and that uh proposes or at least suggests a very straightforward solution to the riddle of factuality because if in so far as you assume that our uh our sensory perception is vertical or at least that uh um that it there's a an overall um uh reliability to our to most of our sensory perceptions. Then uh that reliability simply transfers over to measurement. So in so far as our senses provide factual knowledge, so does measurement. There are different uh types of empiricist views and and different kind of iterations through the history of empiricism on in in thinking about measurement. Uh one uh way of uh of cashing out the um the the factual um uh status of measurement outcomes is through semantic reduction. uh famously Rudolph Carnap in the 1930s um proposed reduction sentences as as ways of um um translating claims about temperature to claims about observable um uh the content of of of claims in obs observation language such as claims about uh the height of the mercury column or the the the number closest to to the um to the to the uh top of the water uh line in a measuring cup. Then there are um other attempts uh conventionalist empiricist accounts. Um they um um agree that measurement does involve various conventions including conventions about what count what count which um intervals of a quantity count is equal or unequal to each other. Um but they still um uh submit that um these human choices about how to measure uh um do not affect the empirical content of these measurement outcomes. So the empirical content of measurement outcomes anything to do with our choices of of units or congruence criteria those um u all those all the human conventions do not affect the empirical content. The empirical content of measurement outcomes is completely uh comes completely from the observations. Um and convention and the conventions merely regulate the way we express those results. Um another u more subtle type of empiricism is uh is representationalism. Um and that's the view that numerical assignments preserve relations among qualitative observations. uh I should say that uh being committed to the uh to the representational theory of measurement as such doesn't make you an empiricist. It's a specific interpretation of um of the axioms of representational measurement theory that I call an evidential uh interpretation that uh that makes an empiricist. Uh that is if you think that um um that it is possible to detect quantitative structure uh simply from the uh from the data itself or through an abstraction from the data without using any theoretical assumptions uh that that would commit you to an empiricist view of the sort that I'm uh critiquing. And to see the problem with these views, uh, we need to take a step back and and distinguish between instrument indications and measurement outcomes. Instrument indications and measurement outcomes are are very different kinds of claims. Uh, an instrument indication, uh, here's an example. Uh, the ammeter needle is between the 1.0 and 1.1 marks on the dial. That's a an an indication of an ammeter. Or uh if you think of a questionnaire such as a such as a or a test such as a reading comprehension test, response option three was selected by the student on item one of the reading comprehension test. That's an instrument indication. It's a property of the instrument in its final state after it interacted with the object or in some cases the person uh that we intended to measure. On the other hand, measurement outcomes are the knowledge claims that we want to infer from instrument indications. These no are no longer claims about the instrument. They're claims uh about the quantity of the or or um the the quantity or more generally measurement value uh that we attribute to the object or person we measuring. For example, the current in the wire is 0.95 uh ampere with a 0.05 uh uncertainty. Uh notice that the [clears throat] number we we uh we assign to a a quantity need not be the numeral that appears on the display of the instrument. That's because uh we may be correcting the instrument for various errors including systematic errors. Uh errors that do not um um average out to to zero in the long run. Uh same with the questionnaire or the or the reading test. The respondent's reading ability is in the 18 80th percentile. That's a an example of a measurement outcome. In this case, uh a norm reference measurement outcome reference to to the distribution of um of uh performance in a some student population. And measurement outcomes are quite different from instrument indication not only in their object that is uh they um they're they they pertain not to the instrument but to the object or or person or event being measured. They're also uh different in that they are they require a specific scale and they uh they involve uh some evaluation of uncertainty either implicitly or explicitly. So now that we know the difference between instrument indications and measurement outcomes, uh let's go back to empiricism and specifically let's ask ourselves what's the relationship between instrument indications and measurement outcomes. Um do instrument indications determine measurement outcomes? And the answer is no. Um um the empirical content of measurement outcomes is underdetermined by instrument indications. Um so it's not just that we could uh transform the scales we use uh by using different units for example. um the the very empirical content of measurement outcomes itself is also underdetermined by instrument indications. And to see that um we we need to think about what happens when we infer measurement outcome from some specific concrete apparatus. Imagine that you're Galileo and you're trying to measure the um the freef fall acceleration uh small g. Um you you rolls balls down inclined planes uh and you measure the their um the time it it it takes them to traverse various uh distances of the inclined planes. Um um from that you want to infer um the the freef fall acceleration. But to do that you first have to make uh several assumptions. You have to model your measuring measurement apparatus um in certain ways. You see above uh a diagram of the way we we we now do it. This is not the way Galileo did it. Um um we we assume uh some idealized scenario. Uh we make some assumptions about um about forces about friction and um and we uh we under the assumption of that model we infer what uh what freef fall acceleration would be. Notice that nothing is actually free falling in our experiment, right? Uh so um so what we're measuring is not uh the real system um as such. We're we're attributing a value to a measure end g uh based on our our model and the observations we make constrain the predictions of this model rather than uh determine >> the value. The value is is is in other words determined jointly by our observations and the model. >> Um can you can you still hear me? Okay. >> Yes. Yes. >> Okay. Um so in other words, additional theoretical and statistical assumptions are required to construct the abstract and idealized model of the measurement process. I didn't mention statistical assumptions. Uh but you can um very well uh imagine that um when we repeat the the measurements uh procedure again and again we we get slightly different results and so we have to make statistical assumption about the distribution of these uh of these indications um over time. For example, we may assume that they're uh they're distributed normally. Uh and if all that's right, then the factual status of observations by itself is insufficient to confer factuality on measurement outcome. Even if we accept that our senses provide us with factual knowledge, uh we we're still relying on the adequacy of our modeling assumptions, the theoretical and statistical assumptions we interpret the indications with. and um and those are not um tested by the um by the measurement procedure. In fact, undetermination is much more pervasive than just what I've uh said. Um and in in my um work over the years, I've shown that not only are measurement outcomes underdetermined by observation, but so are other claims that we also make about measurement. For example, uh coordination claims, those are claims that say instrument x measure measures quantity y. So for example, uh this instrument measures time or this instrument measures uh uh temperature and so on. Even these claims are underdetermined by observation. Quantity individuation claims those are claims that of the sort u instrument X and instrument Y measure the same quantity. Those are also underdetermined by observation. Quantity detection claims, claims about uh the detection of of quantitative structure in our uh data, those are also under determined by observation. And even accuracy claims, claims about how accurate a measurement is >> are under determined by observations. To give you a a taste of some of this, I want to quickly zoom into quantity individuation claims. Um that these are claims about uh um about two instruments measuring the same type of quantity. Uh so here are two instruments. We assume that they're both clocks um and we try to compare their ticks to each other. U so uh imagine that you're using clock uh you you're you're observing the ticks of clock one and you plot them uh um over over the the time um axis using clock two as your standard. And if you use use clock two as your standard, you see uh that clock one um uh has a has a decreasing frequency. It it ticks slower and slower over time relative to clock two. But of course, if you switch the procedure and use clock one as the standard and measure the ticks of clock two, you'll see that clock two uh has a a frequency that's steadily increasing. Now u the question it is perhaps unusual but nevertheless important. Why why think that both of these instruments measure time? In other words, why think that both of these instruments measure the same quantity? Um at least from the point of view of representational theory of measurement, they shouldn't be measuring the same quantity because there is no uh permissible transformation between um their their indications. The the data that we get um uh doesn't have uh the same structure. Uh that's because uh the the intervals of the quantity that clock two measures are ordered differently from the intervals of the quantity the clock two measures. Notice that I'm not talking about uh time itself but rather about time intervals. Right? So the the intervals of of the quantity the first instrument clock one measures are not ordered in the same way. Uh it seems that strictly from the data itself we we shouldn't be compelled to think that these instruments even measure the same quantity at all. But of course we do think that they they're both clocks. um despite their frequency uh drifts. Um we apply corrections to those drifts either by taking clock one or clock two as standard or by taking some other uh um a third clock or um or something like UTC which is a a complex weighted averaging um of of many clocks. We take those as standards and we correct all the other clocks to the the standard. But that uh presupposes that all these instruments measure the same quantity. Uh in other words, the data itself does not force us uh either uh to accept or reject the claim that uh instruments measure the same quantity. Our modeling assumptions are such that we uh assume that they measure the same quantity and then we test whether the consequences of modeling them as measuring the same quantity u mutually coherent. So I hope that's clear how models uh feature in um in claims about one of the individuation from this rather uh uh brief sketch. Uh but let me uh move on to uh the second type of explanation uh causal explanations of the factuality of measurement. You may think um well when we measure we we're interacting with something whether it's length or time or temperature or pressure. Uh measurement is a causal process and um and that causal process uh um in the end has some effect and that effect is the instrument indication. So differences in the magnitude of the measurement cause differences in instrument indications. Uh for example, differences in electric current cause differences in the displacement of the amter needle. So couldn't we simply say that measurement produces factual knowledge because it tracks these causes? Measurement involves reconstructing the cause from its effects. um we observe some effects and from these effects under some assumptions we infer the cause. Uh and if that's the case um then measurement outcomes are factual in so far as they report the detection of a cause. They report the detection of the magnitude um of the measurein that um that was the cause of observed indications. And there are various uh types of of causal accounts out there in the literature. Um um a few books here by uh by Trout by Katrite by Borspoon that take um different um variants of this view. In this uh talk I want to focus on uh on one such account by Luca Mari, Mark Wilson and Andrew Maul that uh in in their book measurement across the sciences that takes a very close look at how metrologists model their measuring instruments. Metrology is the uh the science of measurement. Metrologists are the physicists and engineers who work at at standardization bureaus like the international bureau of weights and measures near Paris and who um calibrate measuring instruments maintain standards such as the standard kilogram and and meter and second and so on. Um, so if anything knows anything about measurement, it's it's these metrologists. They uh um we we better uh look at how they infer measurement outcomes from instrument indications. And um and here's u Mary at Al's um account of how they do it. Um um they say that um the the measure end the thing we're trying to measure is what they call an effective property a cause uh that interacts with our measuring instrument. Um that's in in the diagram. Um that's that's listed as the arrow going into the measurement measuring instrument that says property being measured. That's the cause that we're trying to uh detect. Uh measurement outcomes according to this account are inferred by modeling this causal process. But of course there are other causes intervening causes that also influence the indications of the instrument. Some of these are influence properties. Um think about the um the internal resistance of of the amter itself. or or background magnetic fields. They also affect the the deflection of the amu needle. Uh some of these are affecting properties. These are properties that affect uh the property being measured itself before the the measurement uh measuring instrument even gets into the picture. And um and and modeling the measurement process involves uh making various theoretical assumptions about the causal interactions between all these variables and correct and and and allows correcting uh these various sources of errors. Um what what a um a measurement uh model um provides is um is what's called a calibration function. That's a prediction about the relationship between the measurement outcome the the the quantity we're trying to measure that's listed here as O [snorts] and uh various variables. One is the instrument indication itself or multiple instrument indications a set of instrument indications and all the other eyes I1 I2 I3 are these various influence and affecting properties. Notice however that um the the calibration function here inside the the box this f um is a prediction. Um it's a it's a consequence of modeling the measurement process in a particular way and there are different ways to model the measurement process under various assumptions. In order for causal uh accounts of factual uh status or factuality to get off the ground uh we need to be able to um to arrive at a factual knowledge claim about the the measurement outcome. uh simply a a hypothetical knowledge claim about what the outcome would be under the assumption of this or that model just wouldn't do. We want um factual knowledge claims about what what the temperature is. And that's where uh causal accounts fail. uh to see that uh think of u a simple measuring instrument like the one at the bottom of this slide a caliper. A caliper is supposed to measure the length of uh of some object um between its jaws. Metrologists calibrate uh this caliper by modeling uh its its various uh influence and affecting properties and making various causal assumptions about the relationship between variables affecting the caliper such as temperature uh and and various um um various other variables such as the roughness of the contact and um and the and the abbey error which is the wiggle of various parts of the um of the caliper. Um then they they they provide a prediction of what the um of what the um the length of the object between the caliper jaws would be under those assumptions. But to test those assumptions, they have to use uh some measurement standards such as the gauge blocks uh that are just above the caliper in this uh in this image. These are a metallic uh object of known length. But hold on, how do we know what the length of those objects are? Uh turns out metrologists have to model the procedures by which they um they measure the length lengths of these objects. Uh this usually involves something called a coordinate measuring machine. [snorts] Um and those machines also have to be calibrated uh through a model that predicts uh the um the the dimensions of objects. Um, and I think you're starting to get the picture. Eventually, we get we go all the way to our primary realizations of the meter. Um, those are um um today uh um various uh interferometers based on um agreed upon frequencies of lasers such as this neon laser. Um but even then um the modeling doesn't end. There's no convention uh waiting uh at the end of all this inference that tells you what a meter or or a millimeter or a micrometer is. Uh they're just more and more models. In the end these uh these models are simply compared to each other. The consequences of modeling different in interpherometer lens standard are simply compared to each other uh and shown to cohhere within the expected uncertainties. Um and um so this entire chain of models uh notice that um that the theoretical complexity of of these models only increases with each calibration. There's never a resting point at which we can say that we've detected the cause. We've detected um some independent confirmation of what the um the measurement outcome is. Uh all we get is more and more sophisticated and and supposedly accurate modelbased predictions but no independent confirmation of these predictions. simply uh more and more coherence tests um for these various modelbased predictions and this is what I'm calling an inflation of auxiliaries because with each u with each step more and more assumptions have to be made and and more more sophisticated and and uh um um and and uh strong assumptions about uh causal interaction interactions and about simply uh um um properties um theoretical properties states [clears throat] have to be made um to drive the point home even uh more concretely uh consider the case of the kibble balance. This is um a design of a of a balance that's considered to be uh one of the most if not the most um accurate balance We currently have um here is one such uh balance uh at the national research council in Ottawa in Canada where uh metrologist Carlos Sanchez is um is using it to uh measure the plank constant. The plank constant uh today serves as as the basis for defining the kilogram. Uh um but before 2019 um it it had to be measured extremely accurately uh so that um it's its value could be set um in a way that make made the kilogram after the redefinition uh consistent with the value of the kilogram before the redefinition. Uh before 2019 the kilogram was defined um directly by the mass of a of a specific object. the international prototype of the kilogram. Uh but to give bring you back to this kibble balance, the the idea behind the kibble balance is that it uses uh um uh rather sophisticated quantum mechanical principles uh to to determine the plan constant using um um using the uh using only the the the base uh units of the international system of units, the SI that or also also known as the metric system. That is it determines the plan constant only by using the um the um the definitions of uh the meter um the the second and the kilogram. But to do this, to know how accurate this instrument is, you of course need to model it. Here's what's called an uncertainty budget that um that metologists construct in order to uh um to evaluate how much uncertainty is contributed uh to their measurement by various sources. So uh type A is a a statistical type of uncertainty. Then there are various uh so-called systematic sources of uncertainty such as voltage, resistance, mass, gravity, velocity, alignment and so on. [clears throat] You you you may notice that the uh the uncertainties are uh in parts per billion. So they're extremely small. Uh but to get uncertainty so low, you have to be committed to a vast uh web of theoretical assumptions. Um, not only is u is each row in this uh in this table [clears throat] the result of uh very careful measurement and calculation each row in this table is associated with its own sub budget of uncertainties. Um um for example uh the gra gravity uncertainties involved in in measuring with a kibble balance have to take into account the effect of the earth's tight the the effects of polar motion uh and atmospheric uh pressure. They even have to take into account the effects of the attraction between the mass and the balance itself. Um so in other words, the more accurate your measurement is, the more uh u um uh theoretically uh and statistically com complex your model is. More accuracy means more commitments uh commitments to to more to stronger assumptions uh and to more assumptions about how the instrument works. uh calibration does not free the causal models from dependence on background assumptions. The opposite is true. Calibration deepens the dependence of causal models um on backgrounds presupposition. So again there is no point in this entire inference where um uh where we independently detect uh some mass as the cause of um of our measurement outcome. Um so what do we make of all of this? Um should we simply give up the idea that measurement produces factual knowledge and uh and acquies uh uh to the view that measurement outcomes are consequences that are conditional on model based assumptions. that all we do when we measure is compare um the predictions of various models to each other. In some sense uh the answer is yes. We should acquies uh to this uh to the idea that there's no external uh confirmation for measurement outcomes uh um independently of models. But that doesn't mean we have to um um we have we have to um do without the idea that measurement produces factual knowledge. Rather what we have to do is uh think differently about what factual knowledge is and this is where uh pragmatism uh gets into the picture. According to the pragmatist solution that I'm offering, the factuality of measurement is not due to any special ability of measurement to detect and isolate empirical structures or signals or causes. Uh um rather the presentation of measurement outcomes as categorical factual statements is a pragmatic choice and this choice [clears throat] is justified by the convenience of expressing knowledge claims in a most modular way that is fit for the purpose at hand. So the modularity of knowledge is the ability to take uh some claims and uh repurpose them, reuse them in different contexts an analogously to the way we can use a Lego brick, detach it from one uh structure and attach it to another. Um um so modularity and specifically epistemic modularity is the ability of knowledge claims to quote unquote travel autonomously between context. >> A closely related um idea is the idea of ignorance accordance. Um ignorance of coordinance is that um um the knowledge user uh needs to know little about the context of knowledge production. Think about u uh two factories on one and on you know one in uh Brazil and one in Japan. [clears throat] uh the the the factory in Japan Japan generate produces um uh uh nuts and the and the one in Brazil uh produces bolts or the other way around because now I see the the the image uh is the other way around. What what makes them so confident that the the that in the end of the day the bolt will will fit into the nut? Um you may say well it's the network of uh methological calibrations that trace uncertainties in dimensional measurement uh uh to primary standards and and that's exactly right. What this network affords is uh is ignorance um that is a very specific type of ignorance. the the the knowledge uh users in each of these factories. They don't need to know about the specific assumptions under which the coordinate measuring machines or interferometers or whatever the the other factory used to measure the diameter of their bolts or nuts. They don't need to know about the assumptions that went into these measurements. Um, as long as the entire network is calibrated, they they can simply transmit uh the the dimensions of these objects and be assured that they fit within the uncertainties uh ascribed to these measurements. Um, and the and the conceptual kind of flip involved in in pragmatist thinking um that I'm proposing is that measurement outcomes are not modular because they're factual. Rather, they're are factual because they're are modular. In other words, it's not because we have independent confirmation of the lengths of these or or diameters of these objects that we're we're able to ignore the conditions of of of the productions of knowledge about them. Rather, it's uh because we we manage to produce a a coherent set of uh practices, coherent set of instruments, uh procedures and and theories and models. uh that that um that make our knowledge modular. It's because of all of that that um that we can treat the um the outcomes of our measurement as factual knowledge. Um so you may ask well under what conditions are measurement outcomes modular then um and the answer is that measurement outcomes are modular because they are coherent and secure. Uh so coherence is a combination of nomic coherence coherence among the the laws or or um or functional relations among variables presupposed by models and uh predictive consistency consistency among the the consequences of these models the predictions or retradictions of these models. In other words the reproducibility of results. But coherence by itself is not sufficient. Um because u we want uh we we don't just want um um measurement outcomes that uh that are uh uh context independent. We also want those measurement outcomes to be robust u um in case u of various failures of our background assumptions. And to do that, we also need security. That's the robustness of knowledge claims across a range of epistemically possible scenarios. And here I'm relying heavily on the work uh of Kent Staley uh who who has written about uh this notion of security of evidence claims. >> Um metrologists u dedicate much of their time and effort to securing measurement outcomes. Um so not only are measurement out outcomes uh coherent in the sense of um being um nomically uh uh coherent and and predictively consistent, but also um um met mologists make an go to extreme lengths to to secure the um um the the measurement uncertainties associated ated with different measurement outcomes to trace them to common standards. They use key comparisons by by comparing different um um measure measurement outcomes produced by different labs to each other. Uh they revise uncertainties of incompatible outcomes [clears throat] and uh very importantly they avoid contested assumptions. um uh theories that have not uh uh yet been uh proven to to produce consistent knowledge or or where um physicists uh do not know how to apply a theory well or to to uh um or or to produce um consistent uh predictions from a theory. these uh these theories would not be used when you um uh when you model a measurement process uh in metrology. So only the the most uh secure um and and tested methods are used to to analyze data and to model measurement processes. And that uh that allows measurement outcomes uh to be secure across that is robust across um various uh epistemically possible scenarios, various potential points of failure. If my account is correct, then uh we have good reasons to shift from a truth oriented thinking about uh measurement outcomes to a pragmatist or useoriented thinking about measurement outcomes in so far as we um uh put them uh in a factual mode in a categorical mode. Uh under a truth oriented thinking, factuality is reducible to truth. When we say that um um that the temperature of an object is such and such, we're simply uh claiming that it is true that in some correspondence um sense of truth that uh the temperature of that object is such and such. Under a pragmatist way of thinking, factuality is relative to use, relative to uh my need to communicate the temperature of this uh object to others. Um my claim is modular enough uh for me to treat it as a fact. Under a truth oriented thinking, measurement aims to discover facts, not to generate hypothesis. But under a pragmatist uh solution, measurement is a form of datadriven modeling. Uh it generates modelbased predictions and there is uh uh there's no requirement for it to uh to go beyond or kind of externally to any model. Uh under a truth oriented thinking, uh modular knowledge behaves at best as if it's factual, right? there's this as if uh clause, but under a pragmatist way of thinking, scientific facts just are the most modular knowledge claims available for a given purpose. There's uh there's nothing more that we can or should ask of measurement. Um and that I believe uh solves the riddle. Um measurement outcomes are factual after all. Even though models underlying measurement cannot be independently tested. Even though all the knowledge that we get from measurement is conditional or on on this or another model. U um measurement outcomes are factual because of their modularity because of their um coherence and security. uh measurement outcomes can be used as evidence to test hypothesis because the because measurement outcomes are usually more coherent and secure than um than the hypothesis they're used to test. [cough] So this uh this way of thinking explains why for example we still have u justification for using measurement outcomes uh to test the predictions of uh computer simulations or um of various u um numerical or statistical uh models such as uh machine learning uh datadriven predictions. Um in in the vast majority of cases, the models uh the the models that underly the measurement process and that based that that that underwrite the measurement outcome. Those models are far more coherent and secure than the models that uh that underly the simulation or the or the numerical method or the uh datadriven machine learning prediction. In other words, uh the difference between the those two types of predictions is not uh is not in kind but rather in um um it's it's a it's a quantity a quantitative uh uh um uh difference in coherence and security. measurement finally need not be evidentially superior to other forms of datadriven modeling um such as theoretical prediction and computer simulation. So this uh an interesting consequence of the pragmatist view is that we could at least in principle generate theoretical predictions that are just as coherent and secure as me as measurement outcomes. And in that case we could use theoretical predictions to test our measurements. Um and um there are cases when where um arguably this is already happening that I'm happy to uh discuss but for now I want to thank you for your time and attention. I very much look forward to the discussion and here are a few references from this talk. Thank you very much. Thank you for this wonderful talk. Now as usual we will have five minutes break and then we will back here maybe find a way for him to see us. >> Yeah work on that. >> Yes. And uh yeah, back shortly. [music] >> [music] >> Hey. [music] Hey. Hey. >> [music] [music] >> Hey. Hey. >> [music] [music] >> Hey, [music] hey, hey. [music] Hey. Hey. Hey. [music] Ah, [music] hey. [music] >> [music] [music] >> Heat. Hey. Hey. Hey. [music] Hey. Hey. Hey. [music] [music] >> [music] [music] >> Heat. Heat. N. [music] >> [music] [music] >> Hey. Hey. [music] >> [music] [music] >> Maybe. Okay. Yeah. >> Okay. We are back. So, time for questions, comments, observations. Hi, thank you very much for the talk. Really appreciate very clear. Uh I was I was thinking why in your pragmatic approach you you kept factuality because my impression is that you only need epistemic authority. You say they have they have epistemic authority because they are modular blah blah blah. Why do you need factuality? our microphone are off maybe. >> No, I'm I can hear you. >> Okay, thank you. >> Why kept why keep factuality in a parameist approach? >> Yeah. So, um the I mean the the short answer is that we don't really need factuality uh as such. Um um if um if you're already willing to accept that evidential power, the evidential power of measurement doesn't require factuality. That all that evidential power comes down to is uh something like the degree of confirmation or degree of uh of confidence um or justified confidence in a claim. and that uh and that you're convinced by my account that we have a higher degree of justified confidence in measurement outcomes than we do for example uh in in the in consequences of of most uh scientific computer simulations and and and the consequences of most numerical methods or um or datadriven uh machine learning predictions. Uh if you accept that then we don't need factuality to establish evidential power. The the there are two issues um um be behind the appeal to factuality. [clears throat] One is simply recovering an intuition, right? uh there is a broad intuition uh that perhaps you personally don't share but I think many people do that measurement does provide us with knowledge about u what the temperature is not just what the temperature would be um so one of the goals of the pragmatist solution is simply to recover that intuition to explain why we're able to keep going uh move around in the world as if measurement outcomes to provide factual knowledge um and still succeed in our everyday dealings in in uh um and and not just our everyday dealings with with measurement but also scientific dealings. How how can scientists get by with reporting measurement outcomes [clears throat] as factual claims? If you open a scientific journal, you'll see that measurement outcomes are usually reported as you know the mass of this molecule is not the mass of the molecule would be blah blah under assumptions XY Z right measurement outcomes are are um are reported as factual claims. So one function of the primitist account is simply to explain why we can get by and why scientists can manage uh successfully and safely uh to treat measurement outcomes as factual knowledge despite all that we know about how they're produced. So that's one function. The other function [clears throat] uh of the prognist account is to recover the directionality of testing. We usually use measurement outcomes to test theoretical predictions. We don't usually use theoretical predictions to test measure measurement outcomes. I I I'm acknowledging the term usually here because there are cases where we could be using theoretical predictions to test measurement outcomes uh justifiably, but usually it's the other way around. Uh and the prognistic account uses factuality in order to um um in order to recover that uh that idea that that testing is directional even though uh measurement outcomes are um strictly speaking hypothetical claims. Is that is that clear? >> Can I help? Okay. Thank you. It's very clear but but but I think it's dangerous. It's dangerous because of course the usually is important here. There's cases where I would trust more a simulation in certain case where it's very difficult to measure and if we talk about authority or credence we can compare but if you talk about factuality automatically uh people would say anything factual is better than something non-factual. So it's why it's why it's why I I I in a pragmatic approach I find the factuality problematic because you're you're losing this aspect that sometimes there will be more more authority for a simulation in the case where data are difficult to get. In those cases, I would say that uh the the simulation provides factual knowledge or that it provides knowledge that is at least as factual as the measurement. Factuality in the pragmatist account uh is is not a a um is not a binary variable. It's a continuous [clears throat] variable. It's a degree of coherence and security rather than a a yes no um type property. Uh I I I see the the type of uh of danger that you're alluding to. Um um and um and I share the intuition that um that um it would be in a in a way simpler if we simply you know just gave up factuality altogether. Um but I think um it's it is important to re to explain why we're able to uh to use measurement outcomes uh as if there they are factual claims and and I think it's a um it's a it's a advantage rather than limitation of the account if we can also explain why is it that we can use some the results of some computer simulations as factual Right. Uh while we can treat the results of some computer simulations as factual perhaps is as more factual than the results of measurement outcomes. Uh I think that's an advantage of the prognies account rather than a disadvantage of it. >> Okay. Thank you. >> Question. >> Yeah. I've got I've got so hello this come but this is actually coming in from online uh not from me but I've got uh Judian here who says is joining us live on YouTube from Edinburgh so uh hi Jun [snorts] uh Judian says many thanks uh Erin for your very interesting talk I have two questions I'll give you I'll we'll do them one at a time uh first how do these pragmatic choices of what counts as factual that you talk about today relate to the legislative activities in measurement that you talked talked about in your 2016 uh making time paper. Are you is that a are you shifting from a a more transcendental approach to a more pragmatic approach or how do you how do you see those relate? >> Yeah, that's a great question. I'm still trying to figure that out myself. So, by the way, Julian is the person who actually invited me to give this talk. So, special thanks to to Julian for >> He is there. He saw the talk. So, >> yeah. No, I I I know, but I'm responding directly to a question from him. So, it's [clears throat] um I I want to acknowledge uh um his his efforts in making this happen. Um so, um I'm this is something I'm still um I'm still figuring out. So, I'm very glad u to have the opportunity to kind of think think out loud and and see what uh what the others uh think. But uh um to to give you some background uh in my 2016 article making time, I talk about um the legislative choices that metologists make um regarding the which realizations which concrete processes count as better or worse realizations uh um of uh a given uh definition. >> [snorts] >> uh and those choices are different than the the the choices of the definition uh itself. So uh um it's it's long been noted that part of measuring involves a choice of um of what's called congruence criteria or uniformity criteria. criteria for what which intervals of a quantity count as equal. Such as uh choices about which um which processes are are um have a uniform frequency, right? Is the rotation of the earth more uniform or is the uh the frequency associated associated with atomic transitions more uh more uniform? Today we we go with the latter. For many uh centuries we we went with the former. That's a choice. Uh but legislative choices have to do with uh which which concrete clocks u uh we use to um to approximate the uh the definition with uh even if we define a a uniformity uniform frequency based on um an atomic transition. We still have to ask ourselves which of these concrete objects, concrete clocks ticking away in various labs are closest to uh to realizing the definition. The definition itself is ideal. And there I argue that we have an additional uh choice to make. Metrologists have an additional choice to make. And I call that choice legislative. Uh now the uh legislative choices are not uh are not uh completely conventional in the sense that nature [clears throat] does push back uh different choices uh of um of concrete procedures, concrete instruments would uh would result in in different um in in different uh properties of our measure. Our measure of time, for example, our measure of length or or mass u um um our measurements would would be more or less stable and and more more or less accurate or involve more or less uncertainty based on uh the choices we make. Um uh but there is also a a pragmatic aspect to these choices um in that um we uh we may prefer some of them um on on the basis of um of convenience on the basis of of the utility or even on on the basis of uh political and social uh constraints. For example, [clears throat] the International Bureau of Rights and Measures uh tries to include clocks from many different countries. And that's partially because uh all these countries pay membership fees to the International Bureau of Weights and Measures and uh um and and so there's an an economic uh reason to include all of them. Um so u I see pragmatism as already uh built into the idea of of these legislative choices. Um uh there are there are practical reasons to choose this or that uh clock say um as as a um or or or practical reasons to distribute weights among clocks in a weighted average in the way that say the international bureau of Rights and Measures does. Um, but uh there's a there's a second half to to Julian's question which has to do with the transcendental aspect which I which I know is of interest to many other group members here. Um, [clears throat] I haven't I haven't mentioned uh the the transcendental uh inspirations for this work. Uh partially because I'm not quite sure that uh that the word transcendental really fits into what I'm doing. But uh uh there is at least a loose inspiration uh in the way I use models and modeling in my work on measurement. [clears throat] I use uh uh the generally the uh the requirement to model your measurement process [clears throat] and the the specification of a theoretical statistical model of a measurement process is a necessary condition for the possibility of arriving at a measurement outcome. So in the very broad sense of specifying necessary uh preconditions for the possibility of measurement, the specification of of theoretical statistical model is a kind of transcendental uh requirement. It's [clears throat] not a very strong transcendental requirement in in in that it's nothing like you know can categories and I would say it's it's not even as strong as something like a a relativiz or constitutive a priori as as some uh some philosophers um u argue for those in other context. Um but um um if if if we need to translate this into canon uh I would say that the legislative choices that neologists u um um make are uh closer to regulative ideals. They are they they regulate the distribution of errors among various clocks say or different thermometers or different balances in a way that uh allows the entire networks of standards to maintain security and coherence while also uh balancing that with a with a with a more straightforwardly practical social uh uh even economic uh interests. Uh I don't think that really u satisfies the even the minimal u um um u criteria for a condition to be a transcendental condition. But I'm happy to open this discussion and and hear from others including Julian uh about what what is there something still tension transcendental about my my model based uh approach after all this? I'm I'm curious to know what you think. >> Second question if you want to respond to that. Yes, there is somebody who wants to comment this and find out. Why don't we let people think about that and I'll while I do I'll do Judian's second question and we can yeah so had one more he says uh he asked what do you make of Alistister Isaac's work on high precision measurement particularly of constants he claims that high precision measurement can escape a merely coherentist picture because precision understood as reduction of random error can be assessed statistically and independently of substantive physical theory. It reveals a quote theory independent kernel in the result. >> Yeah, I I've had many constructive disagreements with Alistister over the years and I continue to really enjoy uh our exchanges. Um I I disagree with with Alistister on the on this specific uh uh point, but I I first want to highlight the the great value of of the work that Alistister is doing. Um because [clears throat] um um I think I think he's highlighting a very important counterintuition um a a type of realism that is different from the this the kind of standard uh structural realism about uh entities um um that um sorry structural realism that about relations uh that is is meant to uh uh subvert uh realism about entities that that we meet in the in the realism literature. Usually structural realism um um u talks about the relations among uh uh say the ratios of different variables in a um in a in a theoretical law and and shows us that these relations uh uh for example um uh persevere uh and and and are approximated across cost theory change or that um these these ratios are somehow invariant. Um but all this the kind of standard structural realism story um um appeals to appeals to theories appeals to laws as they are formulated in theories. What Alistair is doing is he's saying put the theories aside for a moment. If we if we simply look at the measurements, right? Um um and and we ask ourselves um when we measure say the uh the the proton electron mass ratio, right? [snorts] these numbers um over years and years of of measuring them converge to uh uh to to just a certain a certain ratio, right? Um and yes, there are broad theoretical commitments involved. Alistair isn't denying that. But there's a convergence of these numbers that is independent of the specific commitments of of very you know very high level sophisticated theories like the standard model of physics right so we have reasons to be realists according to Alistister about uh about about some of these constants um independently of more of of of of more advanced theoretical uh considerations. Uh so first of all I think this this argument has a lot of uh force but um um but where I um where I I diverge from Alistair is really on the um on the spec specific question of what this uh convergence on on um on quantity ratios uh um means what it teaches us. How how far uh um does it establish um um or put it differently, whether or not these these convergences can serve as a foundations of knowledge from which all other measurement outcomes can somehow uh um get their factuality, right? Whether in other words um these um these stable uh fundamental uh constant measurements provide something like an independent uh calibration of all our measurements um independent of any model of a specific of any specific measuring instrument. Uh this is where we disagree. Um all these measurements including the the proton electron mass ratio uh and the fine structure constant um all of these measurements um presuppose first of all certain um conventions of uniformity or congruence the ones that I mentioned in my response to Julian's first question and those are are uh are those are necessary preconditions for these constants to even being constant, right? If we if we measured time or mass with different congruence criteria, uh some constants would just disappear from from our laws. They wouldn't those constants wouldn't be there. And that's simply because uh um the the mathematical shape of our laws would be different such that the quantities in questions would in question would no longer have a a constant ratio. Um in addition to those congruence or uniformity criteria, there are more idiosyncratic uh specific assumptions about the distribution of measurement errors uh including systematic measurement errors across our different instruments. Um um it's it's it's precisely because metrologists uh distribute measurement errors the way that they do across different instruments that we manage to get this convergence that um that Alistister takes to be a sign of uh of some independent confirmation. To me, that's evidence that um um that um that mologists make certain choices that allow uh simple um um stable constants to emerge from their measurements. Right? But it's it's the models of the measuring instruments that are doing the heavy lifting in allowing us to uh to discover uh stable relations. Uh for Alistister, it's the opposite. It's the stable relations that provide evidence that uh that that measurements are getting at something external, some some stable properties in the world. Um and and and and in in a way this is really a um a kind of a rehearsal of of uh arguments um um for scientific realism versus arguments for kind of um a kind of uh cautious um um agnosticism about about um measure and values. um that that is similar to to other debate older debates around scientific realism that that I'm sure you're familiar with. So that's a very long answer, but I I hope it it at least clarifies how much respect and and admiration I have for Alistair's work and at the same time where we disagree. Thank you. Uh there are other questions on uh now that it's the question period. Could you talk about the information also account time just a little bit the flavor? >> Yeah. Yeah. Um so so an information theoretic account of measurement would say that measuring instruments transmit information in some way right uh and in a in a standard kind of Shannon Weaverbased uh uh uh information theory sometimes called syntactic information theory uh we would model a measuring measurement uh measuring instrument like a communication channel where we have some signal coming in from the side of the world. The signal being say the mass of the object or the um or or or the or the temperature of an object or so on. That's that's on on Alice's side, right? The a Alice transmits. And then Bob is us on the other side of the instrument registering um uh registering its indication and uh and information transmission uh is successful uh to the extent that the instrument is sensitive that is u um differences in the u in the indication say the the if we're thinking of an ammeter differences in the um uh in the deflection of the amter needle um uh track the the signal uh that is the current the electric current uh uh entering the ammeter. Um in in some ways this picture um is helpful. Um um there are for example it's useful uh to model um uh certain um certain um aspects of measurement such as uh the level of noise involved in measuring. Right? Some um some sources of uncertainty in measurement uh behave like uh like limitations in in channel capacity, right? They behave like um like interference in in signal transmission and they can be modeled with the same equations. Um um but the problems start when you try to use an information theoretic account of measurement [clears throat] to ground the factuality of measurement outcome to to ground the the claim that measurement outcome is simply a kind of signal detection exercise. Um and when you try to do that important disanalogies uh start to arise between the information theoretic story and the measurement story. One really big difference is that when we have communication, we have independent access to Alice's message, right? We can independently ask Alice what did what what was the message you tried to transmit to Bob and then compare it to the message Bob received and see the differences see the errors. We don't have that with measurement. We don't have any independent access to the current or temperature or the mass uh um of the object we're trying to measure independently of some other measuring device. uh at least not with the quantitative exactness that we need in order to detect errors, right? We may have very crude sensory uh um uh perceptions of which masses are, you know, which objects are heavier or which objects are harder, but not to the not with the exactness that that we need in order to to um uh to to measure um information um uh transmission or information quality. Um that's one problem. But there there's a there are other problems uh other disanalogy between the information theoretic story and the measurement story. Uh another danalogy is that u um measurement doesn't involve transmission but rather transduction. This is a point that Lucamari makes in many of his works. >> [clears throat] >> What we're trying to do when we measure is not um is not reconstruct uh the the signal in the origin but rather convert that signal transduce it to some other property. Say we we transduce um temperature into the volume of mercury or we transduce current into an angle the angle of amter displacement. A third disanalogy and this is where I think things get really uh difficult for the information theoretic story is that many of the uncertainties involved in measurement just don't come from the the procedure at all. Right? They don't come from the the the interaction uh that supposedly transmits the information. uh if you go back recall the those many tables of uncertainties the uncertainty budgets of the kibble balance that I presented earlier in the today. Um uh these uncertainties uh some of them have to do with physical interference like magnetic fields interfering with the with the uh um with the motion of the the balance or with the um um uh or with the with the current running in the coil. But some of these uncertainties have to do with background assumptions. Uncertainties about the evaluation of uh fundamental constants that are being used in the calculation of the measurement outcome. Uncertainties that have to do with the calibration of various standards that were used in the calculation of the the measurement outcome. um uncertainties that have to do with various error corrections um uh corrections for say gravitational potential the earth's gravitational potential um um or um or with with uh with with other um um um kind of statistical um error um statistical modeling techniques um um uncertainties involved in the statistical approximation of of of uh various um parameters that are um that are idealized and therefore require uh some inference. All these uncertainties um cannot be modeled as sources of noise. they have to do with the inference from the observation uh of the measurement process to uh the the value of the measurement. And here you don't really have a good information theoretic account um at least so far we don't really have a a good information theoretic account uh of of how to uh to think of those as uncertainties related to signal transmission. Thank you. Other questions? >> Yes. Um, thank you very much for the talk. Um, I wanted to come back to the actually to the one of the first questions we started discussing. You mentioned that you have some cases when you sort of instead of verifying or testing a theory of hypothesis against measurement you do the vice versa. Can you talk more about this examples please? >> Yeah good. So um um um there there there are several so there's an interesting uh first of all philosophical debate uh going going on for at least a decade uh or more now um um about the epistemology of computer simulation and about uh whether some computer simulations um can produce uh evidence or perhaps even be used as measuring instruments. My my own uh former doctoral supervisor Margie Morrison uh argued that under some under some conditions computer simulations um can provide evidence that is on a par with measuring instruments. Um um there's a um a very nice u article by um Wendy Parker on data simulation that argues that the um the results of some computer simulations in in climate modeling um can be used as uh as measurement outcomes and in fact are used uh very much like measurement outcomes. Um uh so to uh perhaps perhaps to use uh we can we can start with her example Wendy Parker's example from from the data simulation paper. Um uh she describes um uh [clears throat] computer simulation uh computer simulations of past weather. Um [clears throat] and um and these computer simulations are used to infer say the temperature of the ocean in a place where there's no thermometer. And it uses it does that by by simulating the the dynamics of of the oceans, the air, the the the you know the uh solar radiation and so on. uh um in conjunction with data measurement data uh from other points in the ocean. Uh this is extremely important because when you want to try to predict future weather or even future climate, you want to have a regular grid where uh where there's a thermometer every say 3 kilometers, right? But in in reality, we don't have a thermometer every 3 kilometers in the ocean. So these data simulation simulation results are then being fed into forward-looking climate uh predictions and treated as if they are the initial conditions of the simulation. So they're treated as if they are measurements. Wendy Parker argues that there are good reasons uh to to think of those uh as as measurement in as much as um all measurement is model based and uh there is no um um the the the basic requirements for for uh for measurement. namely the the ability to calibrate a measuring instrument against uh known quantity values and use the theoretical statistical models to extrapolate or interpolate from those values to unknown values. That basic requirement is met um um in in the case of uh of these computer simulations. I think that's an interesting argument and I and I think that um um um that it's a it's a case where uh where computer simulations where there's a where there's a strong case to think of computer simulations as measuring uh uh instruments. Now, that doesn't completely answer your question yet because the question is um what if we had a thermometer there that showed a different temperature than the computer simulation? >> Wouldn't wouldn't we trust the thermometer more than we do the computer simulation? And in that point I would agree that yes we haven't flipped the directionality of confirmation in that case because we the the background knowledge for for thermometry is still more coherent and secure than the background knowledge involved in these computer assimulation simulations. Where things start to get more tricky is when you uh is when you look at um um the um the way some uh some measurement standards are themselves um based on uh computer simulation. So if you look at uh thermometry again, but you go all the way up to the most accurate realizations of the Kelvin scale um especially to acoustic gas thermometers. [clears throat] A big part of the of the uncertainty budget um of an acoustic uh gas thermometer which serves as a standards for calibrating real thermometers comes from uh a numerical simulation of the gas that that uh that calculates the uh the speed of sound in the in in the gas uh that is used for for the uh uh for the standard. Um so already some of the uncertainties involved in calibrating a physical traditional thermometer to the Kelvin scale that inference already goes through a computer simulation which I think is very interesting. Um um going on with a series of examples. Um there are cases in in chemistry in computational chemistry [clears throat] where um um um chem computational um uh chemists are allowed are are able to predict with a with very low uncertainties what the say the bond angles of a given molecule um are. Um and um and >> and some of these are based on uncertainties that that are starting to look like they're smaller than the uncertainties involved in experimental methods. >> Um >> so I'm I'm I'm hearing I'm hearing some background. Uh uh is that >> was noise but can can you please continue that that's with this example of chemists >> there's people outside there it's people outside >> yeah okay so so yeah so the the the computational uh these computational chemistry um uh simulations allow you are in fact tracked by NIST by the international standards of of um science and technology in the in the US there's a whole database uh for computational chemistry that NIST uh maintains and that um gradually you see more and more of the uncertainties reported by these computations uh shrinking below the uncertainties associated with measurement and it's no longer clear at least when the measurement and the and the um the the theory sort of abinio or or or semi-empirical uh calculation when when those uh differ. It's not immediately clear which one is at fault. Um um so I think um even even though we don't yet have u um um a computer simulation that um that is a clear u that is is is clearly superior to uh to the best measurement uh possible of a quantity in every in each and every case. We definitely have uh now numerical methods that are superior to many uh to to many uh kind of ordinary not uh not perhaps the best measurement instrument measuring results. >> Other questions? So I will ask one um yes can you say something more about the concept of modularity you were talking about I think in slide 20 and the idea of autonomy that is introduced in reference to modularity. >> Oh sorry you're talking about sorry modularity. Yes. >> Yeah. Um [clears throat] so modular the modularity of of u of measurement outcomes is something we we take for granted um in our both in everyday life and in science. Um, you uh you you you move into a new uh apartment and you want to know whether the sofa from from the IKEA catalog is going to fit in your room and you measure the room and then you compare it to the measurement of the of the of the length of the sofa in the IKEA catalog and you take it for granted, right? We usually take it for granted that the comparison of these two numbers is gonna is going to give us vertitical uh information uh about whether the sofa will fit whether the sofa is going to be too too large or just or you know or small enough to fit alongside the wall. Um and notice that what what we do when we assume that is we assume that the the the procedure and the underlying assumptions of the procedure of measuring my the the length of my wall. Um and and the underlying assumptions that were involved in measuring the length of the sofa by whoever you know somewhere someone in IKEA who who uh who designed it and and constructed it that those uh those assumptions don't really matter uh for for our communication that we can simply ignore the question of you know whether did did they use a ruler did they use uh uh uh you know what kind of ruler did they use? Did they use a um um a measuring tape? Uh or perhaps did they use an an interferometer? Um we don't care, right? We simply don't care what the h how the IKEA uh um uh uh um employee uh measured the the length of the sofa. [clears throat] We we simply compare the results. And this is this is a mundane example, but of course we industry and commerce relies on it all the time. We uh we we can have a a car a car manufacturing factory uh in Belgium uh that and that gets uh it's you know car doors from a a factory a factory in Japan yet they fit with the you know with the rest of the car perfectly um for the for the for the very same reasons that the measurements are deemed to travel between these contexts um allowing us um to completely ignore how those measurements were produced. The the important bit of information that we need is the uncertainty. We need to know what the uncertainty is of the of that measurement in order to know what wiggle room we have. How much uh we can ex expect comparisons between numbers to be informative and at what point uh small differences in number are no longer uh um a good basis for decision- making. So that's what I mean by modularity. uh the modularity, epistemic modularity really is >> a kind of context insensitivity of knowledge. And notice that this is in this is in tension with some uh with with some accounts uh of um of of the epistemology of data that that say that data only travel uh because we have metadata because we have information about the provenence of that of that data. Um and I'm referring especially to Sabina Leonelli. I don't really have a direct disagreement with with that view. uh but but rather I want to uh highlight the the source of the difference. Uh Sabina Lenelli's examples come from uh areas in in the life sciences uh such as genetics where uh where um data isn't modular where uh um um there is there is no way to completely decontextualize the data and simply remove it from one context. say one genetics lab produces information about it some some you know genetic sequence and then just transfers it to another lab. What what Lenelli shows is that [clears throat] the issue is much more complex and and in some cases it is if especially if you want to uh reuse data in ways that um that were not intended uh in in the way that it was originally produced. You cannot simply ignore the way the data was produced. Um it's definitely true in in cases of measurement in uh in in areas like mental health for example measuring measuring uh the severity of depression in one context and I've shown this in other work uh doesn't allow you to make inferences um about the severity of mental health using the same questionnaire in in another context using another population with with a with a different type of community, a different type of say health service and so on. So I'm not claiming that everything that goes by the name of measurement uh is in fact modular in the in the way I suggest but rather I'm pointing to a very specific uh very high standard of of success that is a methological uh standard of success uh methological reproducibility u is an extremely high standard. Um, it's it's built uh um with with a level of rigor that supports ignorance affordance that affords ignorance u of the sort that allows you to fit this the IKEA sofa into your room. Um and that is the high level of uh uh the high high level of rigor and of the high level of uh context um invariant reproducibility that I'm interested in for this talk. I'm not claiming that all uh that everything that goes by the name measurement um um fulfills that standard. I hope I hope that's a little helpful. >> Yeah. Yeah. Yeah. Absolutely. Thank you. >> Um, somebody else? >> No. So, I I also wanted to ask another question. Um we could distinguish different kinds of measurement in the sense that for example we may perform a measurement because we actually want to know the value of something like in the case of um plank constant you were mentioning before uh but we can also perform a measurement because we are directly testing a theory like I don't In the case of 199 19 eclipse when general relativity was tested by the deflection of the uh light of the star behind the the sun uh or we may >> are you taking a bad measurement on purpose? >> What are you taking a bad historical measurement on purpose? Oh, that's an example. >> Imagine it was a good one. Okay. [laughter] >> Yeah. Otherwise, we can have uh other kinds of measurements like uh detection for example when we are were searching for the hig boson at LHC. So uh do we need some kind of differentiation between all these cases for what concerns your perspectiveist sorry pardon meist approach um or not? >> Uh I don't think so. No, I I think in both of these cases when we want to just know the number, we want to know the plank constant or even we want to know the length of the sofa, uh or when we want to test uh a theory such as general relativity or the standard model. Uh the epistemology of measurement is the same in both cases. Um in all of these cases we we would be constructing a model an idealized and abstract model of the measurement process um from theoretical and statistical assumptions. um um test that that model during calibration and then make a prediction a calibration function about uh what the the value of the quantity uh would be. um um given the an indication or or a set of indications that are produced to the instrument and then using that instrument to produce the indication and deriving the the outcome. Uh from the point of view of the the inferential structure of measurement it would be the same. The difference has to do with uh with the reasons we're using it and and and the some of the decisions uh we'll make. So for example, the decision to accept or reject a a theory um [clears throat] um would be um would would involve further considerations that usually go beyond the measurement itself. It would we would for example have to compare the predictions of the theory when and and and those predictions would also have some uncertainty associated with them. We'd have to uh uh to use um certain statistical methods to see whether the the predictions uh agree. Um um but um the the lesson uh that so there there okay I want to distinguish two things first of all u uh there are cases when we measure just for measurement's sake um Ian hacking uh used to call it a measurement fetish there's a fetish for numbers um Um um um I don't want to um kind of exaggerate how much of measurement uh is is driven by simply the the the need to get to the next decimal point. But certainly some um uh some of the drive um to uh for for increased accuracy is simply self- sustaining. We we we want to um we want to get to the next decimal point of the say the fine structure constant or or uh or or other constants um simply because we it's a technological challenge. We're trying to uh increase the scope and the accuracy of our uh of our theories in a way that is is very much reminiscent of what called normal science. Um but that is of course not the only reason why we want that that next digit. We want that next uh decimal digit because we also um there's a chance that we'll discover a new physics at that decimal dig digit that new effects that were masked by the low uncertainties of older measurements that that will be revealed in the next uh digit. For example, when you go from microwave uh uh uh based atomic lock like cesium fountains to optical atomic locks um your atomic locks become very very sensitive uh uh gravimeters. So a difference of a few centimeters in the height of your desk, right? the height at at of of the of the desk that or or the table that holds the clock is going to make a discernable difference in its frequency. That's how um um that that's how um accurate they are. They're extremely sensitive to to the tiniest fluctuation of gravitational potential. Um and that that that means that with every um every push for accuracy, we we can we have the opportunity to discover um testable differences in the test and and deviations from the testable consequences of some of some theory. Um so the two goals are connected. the goal for of just wanting to know the number and the goal of wanting to discover new physics to or or to to reject potentially uh certain predictions of our models. Uh those those are ultimately um uh connected in terms of the goals of metrology. Um it's just that theory testing involves much more than simply measuring. It involves additional inferences uh that have to do with the with a with a derivation of predictions of from theory and and the the comparisons of those predictions with the predictions of a model of the measurement process. In the end, we're we're just comparing predictions though. That is a point I want you to take from this talk that in the end when we test the theory against measurement, we're simply comparing the predictions of the theory with the predictions of the model of the measuring instrument. Um and the reason then we re the reason we we uh we we uh we we decide for against the theory rather than for against the measurement is that the measurement uh the background knowledge informing the measurement is more coherent and secure. [clears throat] >> Thank you. Okay. So I think that we can call it a dealing and we can thank our speaker again. [music]