Looking at the specific features of experiments within the overall practice of science, there is one feature that stands out. In order to perform experiments, whether they are large-scale or small-scale, experimenters have to intervene actively in the material world; moreover, in doing so they produce all kinds of new objects, substances, phenomena and processes. More precisely, experimentation involves the material realization of the experimental system (that is to say, the object(s) of study, the apparatus, and their interaction) as well as an active intervention in the environment of this system. In this respect, experiment contrasts with theory even if theoretical work is always attended with material acts (such as the typing or writing down of a mathematical formula). Hence, a central issue for a philosophy of experiment is the question of the nature of experimental intervention and production, and their philosophical implications. To be sure, at times scientists devise and discuss so-called thought experiments [19]. However, such 'experiments'--in which the crucial aspect of intervention and production is missing--are better conceived as not being experiments at all but rather as particular types of theoretical argument, which may or may not be materially realizable in experimental practice.
Clearly, not just any kind of intervention in the material world counts as a scientific experiment. Quite generally, one may say that successful experiments require, at least, a certain stability and reproducibility, and meeting this requirement presupposes a measure of control of the experimental system and its environment as well as a measure of discipline of the experimenters and the other people involved in realizing the experiment.
Experimenters employ a variety of strategies for producing stable and reproducible experiments (see, e.g., [20, 21] and [6]). One such strategy is to attempt to realize 'pure cases' of experimental effects. For example, in some early electromagnetic experiments carried out in the 1820s, André Ampère investigated the interaction between an electric current and a freely suspended magnetic needle [22]. He systematically varied a number of factors of his experimental system and examined whether or not they were relevant, that is to say, whether or not they had a destabilizing impact on the experimental process.
Furthermore, realizing a stable object-apparatus system requires knowledge and control of the (actual and potential) interactions between this system and its environment. Depending on the aim and design of the experiment, specific interactions may be necessary (and hence required), allowed (but irrelevant), or forbidden (because disturbing). Thus, in his experiments on electromagnetism, Ampère anticipated a potential disturbance exerted by the magnetism of the earth. In response, he designed his experiment in such a way that terrestrial magnetism constituted an allowed rather than a forbidden interaction.
A further aspect of experimental stability is implied by the notion of reproducibility [9]. A successful performance of an experiment by the original experimenter is an achievement that may depend on certain idiosyncratic aspects of a local situation. Yet, a purely local experiment that cannot be carried out in other experimental contexts will, in the end, be unproductive for science. However, since the performance of an experiment is a complex process, no repetition will be strictly identical to the original experiment and many repetitions may be dissimilar in several respects. For this reason, we need to specify what we take or require to be reproducible (for instance, a particular aspect of the experimental process or a certain average over different runs). Furthermore, there is the question of who should be able to reproduce the experiment (for instance, the original experimenter, contemporary scientists, or even any scientist or human being). Investigating these questions leads to different types and ranges of experimental reproducibility, which can be observed to play different roles in experimental practice.
Laboratory experiments in physics, chemistry and molecular biology often allow one to control the objects under investigation to such an extent that the relevant objects in successive experiments may be assumed to be in identical states. Hence, statistical methods are employed primarily to further analyze or process the data (see, for instance, the error-statistical approach by Deborah Mayo [23]). In contrast, in field biology, medicine, psychology and social science, such a strict experimental control is often not feasible. To compensate for this, statistical methods in these areas are used directly to construct groups of experimental subjects that are presumed to possess identical average characteristics. It is only after such groups have been constructed that one can start the investigation of hypotheses about the research subjects. One can phrase this contrast in a different way by saying that in the former group of sciences statistical considerations mostly bear upon linking experimental data and theoretical hypotheses, while in the latter group it is often the case that statistics already play a role at the stage of producing the actual individual data.
The intervention and production aspect of scientific experimentation carries implications for several philosophical questions. A general lesson, already drawn by Bachelard, appears to be this: the intervention and production character of experimentation entails that the actual objects and phenomena themselves are, at least in part, materially realized through human interference. Hence, it is not just the knowledge of experimental objects and phenomena but also their actual existence and occurrence that prove to be dependent on specific, productive interventions by the experimenters. This fact gives rise to a number of important philosophical issues. If experimental objects and phenomena have to be realized through active human intervention, does it still make sense to speak of a 'natural' nature or does one merely deal with artificially produced laboratory worlds? If one does not want to endorse a fully-fledged constructivism, according to which the experimental objects and phenomena are nothing but artificial, human creations, one needs to develop a more differentiated categorization of reality. In this spirit, various authors (e.g., [20, 9]) have argued that an appropriate interpretation of experimental science needs some kind of dispositional concepts, such as powers, potentialities, or tendencies. These human-independent dispositions would then underlie and enable the human construction of particular experimental processes.
A further important question is whether scientists, on the basis of artificial experimental intervention, can acquire knowledge of a human-independent nature. Some philosophers claim that, at least in a number of philosophically significant cases, such 'back inferences' from the artificial laboratory experiments to their natural counterparts can be justified. Another approach accepts the constructed nature of much experimental science, but stresses the fact that its results acquire a certain endurance and autonomy with respect to both the context in which they have been realized in the first place and later developments. In this vein, Davis Baird [24] offers an account of 'objective thing knowledge', the knowledge encapsulated in material things, such as Watson and Crick's material double helix model or the Indicator of Watt and Southern's steam engine.
Another relevant feature of experimental science is the distinction between the working of an apparatus and its theoretical accounts. In actual practice it is often the case that experimental devices work well, even if scientists disagree on how they work. This fact supports the claim that variety and variability at the theoretical level may well go together with a considerable stability at the level of the material realization of experiments. This claim can then be exploited for philosophical purposes, for example to vindicate entity realism [14] or referential realism [8].