Part V — Which Relations around Research Can Move inside an Artificial Configuration?

Chapter 9 — What Relations Surround Research for S?

Roles, Value, Evaluation, and Control

1. Where the attribution of research meets the research system

A research machine may call to mind a single researcher replaced by a single machine. This book, however, has not restricted the one who researches, A, to individuals: A can be a person, a group or organization, or an artificial or hybrid configuration (Chapter 7, Sections 1 and 2). And a system that conducts research may also select its investigators, distribute their effort, criticize their results, and revise the rules by which they work. In human research these have often been held by different parties. For example, universities and laboratories have decided who joins an inquiry; funders and laboratory heads where effort goes; reviewers and other researchers who try to replicate results have done the criticizing; and the communities of a field, among others, have set the rules of work. An artificial system may come to do within itself part or all of what these social arrangements around research have done. Up to Part IV the book has asked chiefly about those who research and their results, and has touched on these arrangements only where they bear on knowledge or on status as a researcher. Yet in human research much of what decides where research goes, such as which inquiries funds are directed to, which results are accepted, and who sets the direction of research, has been settled in such arrangements outside individual human researchers. Part V takes these up as its topic and asks what changes when the relations around research enter an artificial configuration. The question follows from Part IV's account of research actors but reaches beyond an actor's capacity: how its inquiries are related to one another, to the society relative to which their intended achievements would be new, and to those who bear their consequences. These relations are not deduced from the working characterization alone. They arise where what it names as the target of research, epistemic achievements new to S, meets the dependence, plurality, and finite resources of research.

The first reason the question arises is that the achievement named in the characterization is not, at least in the case of knowledge, a property of the actor alone: its novelty is relative to S, and on several accounts its warrant or possession also involves others. The argument here proceeds mainly through that case. John Hardwig argued, as Carolyn McLeod summarizes the argument in her encyclopedia entry on trust, that scientific knowledge depends for its acquisition on trust in the testimony of others, since no single person has the time, intelligence, and experience needed to find out independently what is known (McLeod 2020, §3). On this argument, dependence on others is not a shortcut a more diligent inquirer could dispense with but part of how the knowledge becomes possible; Chapter 11, Sections 5 and 6 return to Hardwig's argument. Chapter 2, Section 6 set out that on several accounts social organization enters the warrant or possession of knowledge itself, and Chapter 3, Section 8 that local warrant does not settle whether a continuing system preserves its capacities for diverse search, independent criticism, memory, correction, and further question formation. In human research, knowledge new to S has therefore often been brought about through relations that extend beyond whatever unit conducts a particular inquiry, and on accounts on which social organization enters warrant or possession, it is also warranted and possessed through them. Whether the first holds where that unit is an artificial configuration that conducts many inquiries is one of the questions that the sorting of Section 6 asks.

The second reason is that the problem of the artificial researcher and the problem of the research system, which correspond to two of the debts that Chapter 1, Section 9 lists as questions the working characterization leaves for later parts to answer, the attribution debt (to what, and within what boundary, research can be attributed) and the research-order debt (questions about the social order surrounding research), cannot be reduced to one another. An account of an artificial researcher and an account of a research system can concern the same arrangement while having different objects. The contrast is not between an individual and a collective: an artificial researcher can itself be a collective or an organization (Chapter 7, Sections 1 and 2). It is between what a researcher, of whatever composition, can do and how inquiry among researchers is organized. Research system here names the latter, the social order of research that Chapter 1, Section 9 calls the research order. That order can be realized in human institutions, in mixed arrangements, or within one artificial system that organizes inquiry among its own instances or investigators, as in the example below. Where it is realized within one artificial system that also conducts research, the same arrangement can be described both as a researcher, a configuration to which the conduct of research is attributed (Chapter 7), and as a research system, an organization of inquiry among its instances or investigators; the contrast is then between these two descriptions, not between two kinds of object. In Part V an artificial research system is an artificial system described in the second way, that is, a configuration considered as the organization of inquiry among its instances or investigators. A population of individually excellent investigators can form a poorly organized community, and a well-organized community can be composed of investigators whom no individual assessment would favor. Conor Mayo-Wilson, Kevin Zollman, and David Danks give one version of this a precise form within a formal model of communities that share evidence. Counting a method as rational when its chance of choosing the best available action approaches certainty as inquiry goes on, they prove that methods that converge in this sense for an inquirer working alone can fail to converge when used together, and that a community can converge although none of its members would alone. They also prove that individual and community rationality are not wholly independent: methods that converge in whatever community they are placed form communities that converge, although even on this standard a community can converge where its members' methods would not. How far individual and community prescriptions come apart therefore depends on the standards of individual and group rationality adopted (Mayo-Wilson, Zollman, and Danks 2011, 653–655, 663–669). Put in the terms of artificial research, how an agent performs when it is assessed alone does not by itself show how it performs among agents with which it shares evidence, and even an assessment in those settings does not settle how inquiry among researchers is to be organized. Nor does the fact that inquiry is well organized by itself establish that any unit, a member or the collective, is an artificial researcher (Chapter 7).

The third reason is specific to artificial research. As noted at the outset, in human research the relations through which inquiry is directed, certified, and governed for S have been realized not only by investigators themselves but also by people and bodies other than the investigator. An artificial configuration can, at least in principle, take some of these relations inside. It can, for example, generate the investigators among whom effort is divided, run the checks through which results are criticized, and hold the decisions through which the whole is directed.

An example, which this book constructs for the purpose and to which the following chapters return, makes the possibility concrete. A single system might generate a hundred instances to pursue different lines, assign other instances to check and reproduce what they produce, and shift computing resources among the lines according to those checks, with no colleague, referee, or funder outside it, although those who built it may still be able to stop, update, or retire it (Table 9.1; Section 5). Parts of such an arrangement already exist: the AI Scientist, for example, includes review among the steps through which it organizes research (Introduction, Section 2; Chapter 11, Section 7). Those who build or deploy such a system can make three kinds of claim about it. They can claim that it allocates inquiry well, moving its resources to the lines that merit them; that the results which pass its internal checks are certified for S, so that others have reason to rely on them; and that it may decide, on behalf of others, which of S's questions its instances pursue, a claim not about how well it allocates but about who may make that decision. The example's role is to show, held in one arrangement, the three relations through which inquiry is directed, certified, and governed, which Section 2.1 names allocation, criticism and certification, and authority. For each relation, Part V asks what it would take to substantiate the corresponding claim.

Chapter 7, Section 2 held that moving a function across a boundary changes the subject of the claim without making the function disappear. That remains true. But it leaves open whether a function performs in the same way once it has moved, and whether every function can move at all. In the example, whether checks run by instances of the same system catch what an outside critic would, and whether a system that shifts resources by its own checks can also settle which of S's questions are worth pursuing, are questions of this kind. The question of this part is therefore:

Which of the relations through which research is directed, certified, and governed for S can be located inside an artificial configuration, and what changes when they are?

Of the two parts of the question, the answer to the first, which relations can be located inside on which accounts, serves as a means of describing the second, what changes in research when they are. The question is not whether artificial configurations can conduct research, which Part IV addressed, and answering it adds no condition to the working characterization. It does bear on questions that Chapter 1, Section 1 left at the level of value and governance, where importance, usefulness, legitimacy, risk, and the like are assessed apart from the conditions of research status and of achievement, and Part V compares the answers that different accounts support and argues what changes when the conditions each account names are realized inside a configuration. Drawing these comparative conclusions does not require selecting one account of value or legitimacy for every decision. Of the questions Chapter 1, Section 1 left at that level, which unknowns merit attention is taken up in Section 3 and Chapter 10, Section 6; whose purposes are to govern inquiry, in Chapter 10, Section 6 and Chapter 12; how epistemic and non-epistemic values interact in certification, in Chapter 11, Section 4; which risks are permissible, as the question of who may set the error costs that others bear, in Chapter 11, Section 4 and Chapter 12, Section 2; and when a result is to alter a decision, only as the threshold at which a result may be relied on for action (Chapter 11, Section 4). Of what Chapter 1 listed at that level, utility, beauty, and emancipation enter only where they bear on which unknowns are pursued, and institutional recognition only where it moves allocation as credit (Chapter 10, Section 3).

The rest of this chapter prepares the question. Section 2 says which relations are meant, who holds them, and what their division among different parties may have supplied. Sections 3 to 5 set out what all three of those relations involve, rather than one relation each: valuing, evaluating, and controlling, in the minimal form that the following chapters use. Section 6 states how Part V proceeds. Some of this preparation already shows relations that a comparison of outputs alone misses: even where an artificial configuration delivers the same knowledge to the same S, on positions such as those that value the formation of researchers the organization of research need not keep the same value, since who took part in the process, and on some accounts of authority who decided, remain in question (Section 3).

2. Three relations and the parties who hold them

Section 1 spoke of the relations through which research is directed, certified, and governed for S. This section asks which relations these are, why Part V counts three, and who holds them. Here and throughout Part V, an actor or inquirer is whatever unit the conduct of research is attributed to: a person, a group, or an artificial or hybrid configuration (Chapter 7, Sections 1 and 2). The section derives each relation from a term of the working characterization that marks where the activity is tied to S, taken together with a condition outside the characterization that explains why that tie becomes a relation extending beyond whatever unit conducts a particular inquiry (Section 1). In human research such relations have been held by parties other than that unit as well as by it (Section 1); whether an artificial configuration that conducts many inquiries can hold them inside itself is the question of Part V.

2.1 Deriving the three relations

The first term is new to S. Chapter 1, Section 7 derived from it, together with attempt, that research has a question structure: the attempt is directed at some respect in S's relation to a target that is not yet settled. It also observed that a society is ignorant of indefinitely many things and that novelty does not decide which of them deserve attention (Chapter 1, Section 8). The term supplies the space within which a choice is made, not the need to make one. The need comes from a condition outside the characterization: the means of inquiry, such as attention, effort, time, instruments, and money, are finite and have alternative uses, so that not every unknown can be pursued at once. Chapter 4, Section 5 and Chapter 5, Section 4.2 argued that the choice of what to take up as a question and what to pursue can be made repeatedly within an inquiry, and Chapter 7, Section 4.3 argued that the capacity to select questions is one dimension of an actor's autonomy. What makes the choice a relation beyond the unit is a second outside condition, that many investigators pursue S's unknowns. Among them the same choice takes a different form: effort is divided among lines of inquiry whose prospects are unknown, and the division is not the sum of each investigator's choices. Call this relation allocation: the distribution of attention, effort, and resources among S's unknowns. Chapter 10, Section 1 takes up why it arises and what makes it hard.

The second term is epistemic achievements, developed here first through knowledge. Chapter 2 separated the production of a result from its becoming knowledge, and set out that on the critical-contextual account, venues for criticism, uptake, public standards, and tempered equality of intellectual authority help constitute objectivity rather than merely communicate it, and that on distributed or group-functional accounts, on which a properly organized collective can itself realize a knower, retrieval, use, and revision can enter possession. On accounts that locate knowledge in individual knowers, the term itself does not point beyond the knower. What points beyond it there is the index to S, the relativity of novelty to S, together with a condition outside the characterization, related to the one Section 1 noted with Hardwig and put here in terms of S: the members of S cannot each gather the evidence for many results, so that their reliance on a result runs through others: through those who produced it and, where it has met scrutiny beyond them, through that scrutiny. For an artificial S, how far this condition holds depends on how evidence is shared among its members, for example whether what one member has gathered is available to the others as their own record or only through its report. Call this relation criticism and certification: the relations through which a candidate result is exposed to error-detecting scrutiny and through which others are given reason to rely on it. How far such reliance has to rest on scrutiny of the result is one of the questions on which the accounts of testimony compared in Chapter 11, Section 5 divide. Certification also covers the judgment that a result is in fact new to S. Chapter 5, Section 5 noted that identifying what is new to S can itself be work within inquiry, and like any such work the judgment can be mistaken. Chapter 11, Section 1 separates the two parts of this relation, exposure to error-detecting scrutiny and the giving of reason to rely. On accounts that locate the value of criticism in the errors it detects, the first part can hold within a single actor, while the second refers to S; on critical-contextual accounts, what a check counts against is itself relative to assumptions a community holds, so that, where that community is S, the first part can refer to S as well.

The characterization, however, names epistemic achievements in general, not only knowledge. The knowledge argument above specifies one important case; it does not establish that every achievement has the same social conditions. For understanding, certification can concern an explanation's adequacy or an attribution of understanding without conferring understanding on its recipient. Claims about an explanation's adequacy or someone's understanding can also be relied on without being checked by each recipient, making criticism and certification relevant here. What certification establishes depends on which claim it assesses. Chapter 11, Sections 5.6, 6, and 9 develop this extension not as an implication of the knowledge accounts just surveyed but as a further argument using the distinctions of Chapter 3.

The third term is S itself. Chapter 8 distinguished the epistemic success of research for S from legitimate governance, and observed that specifying S does not identify who may speak for it. The outside conditions here are that S, short of the limit Chapter 12, Section 5 considers, a single configuration that is both the actor and the whole of S, has many members and no single voice, and that the consequences of inquiry can reach parties who do not belong to S. Which unknowns are pursued and what error costs are accepted can be fixed by someone's decision or by processes that no one sets as policy, as Chapter 10, Section 5 suggests from cases of allocation. Where they are decided, the question is who may decide them and who may revise the decision; where they are not, it is who answers for the setting and who may change it. Call this relation authority: the standing, in the sense defined next, to decide and to revise the terms of the other two relations and the further decisions about inquiry that Chapter 12, Section 2 sets out.

Standing, here and throughout Part V, is meant in a normative sense only. It is an entitlement, held in relation to those subject to a decision, to propose, decide, object, refuse, or revise. It is neither competence, nor the formal assignment of a right to decide, nor the effective capacity to exercise such a right (Section 5). Which parties have standing, and on what basis, depends on the account adopted. Whether competence can ground it is a question on which accounts of legitimate authority divide, and Chapter 12, Section 1 takes it up; whether it rests on features of a being, such as welfare or agency, or on relations among beings, such as power or cooperation, is a question on which accounts of justice divide, and Chapter 12, Section 5 takes it up. Three other matters that the words standing and authority could name are given other terms. Longino's condition of tempered equality concerns what Longino calls intellectual authority, the capacity to contribute to the critical interactions of a community; the condition holds that any perspective has that capacity prima facie (Longino 2025, §5); Part V keeps that term, and Chapter 11, Section 3 takes it up. What an actor's outputs count as for S is called below their status for S. And whether an artificial entity has claims of its own as a being, which Chapter 12, Section 5 asks, is a question of moral status.

The derivation comes with four qualifications. First, the correspondence between terms and relations is not one to one. Novelty to S bears on certification as well as on allocation, and the parties to whom authority may have to answer can extend beyond S to those who bear the consequences of the decisions it takes; which of them it answers to depends on the account of legitimacy adopted (Chapter 12, Section 1). Second, the relations are not on a par. Allocation and certification are first-order relations; authority is second-order in that it sets and revises the terms of the other two. The further decisions it covers, which Chapter 12, Section 2 sets out, include operational and risk decisions, constitutive and recognition decisions about who counts as a member, a critic, or an expert, and decisions about memory and correction. Third, of the three, certification is the relation that can most directly enter a result's status as knowledge for S or its becoming S's knowledge, on the accounts of Chapter 2, Section 6 on which social organization enters warrant or possession, respectively; on accounts that make group knowledge depend on the members with decision-making authority in the relevant domain, a form of authority enters as well. Where a relation does not enter the conditions of the achievement claimed, its assessment remains at the level of value and governance (Chapter 1, Section 9). For knowledge, on the accounts considered here, this is so for allocation, for authority except on the group-authority accounts, and for certification on accounts on which social organization enters neither warrant nor possession. For other achievements, whether these relations enter success depends on their own accounts, such as the accounts of public or collective understanding in Chapter 3, Section 6; on accounts that make understanding a kind of knowledge, the conditions just described for knowledge may carry over (Chapter 3, Section 3). Fourth, none of the three is a condition of research status. In the sense of the working characterization, an activity organized toward epistemic achievements new to S is research whether or not the allocation that set it going was wise and whatever the legitimacy of its direction (Chapter 1, Sections 1 and 9). Part V takes the relations up nonetheless because claims about artificial research systems, of the kinds illustrated by the three claims of Section 1, are claims about these relations.

The division into three is meant to pick out the relations around research whose location inside or outside a configuration bears directly on those claims. Chapter 1, Section 9 took the research order to be made up of the division of labor, the allocation of inquiry, criticism and certification, memory and correction, credit, and authority, and listed agenda setting and attention among what artificial researchers may change. Part V treats those of these that are not themselves the three relations where they bear on the three rather than as relations of their own. The division of labor and attention fall under allocation as defined above (Chapter 10). Agenda setting is taken up under allocation, as the choice of which of S's unknowns to pursue, and under authority, as the question who may make that choice (Chapter 10, Section 6; Chapter 12). Memory and correction are taken up as conditions of certification over time (Chapter 11, Section 8, where memory takes the form of the record) and as decisions under authority (Chapter 12, Sections 2 and 6). Credit was treated as an attribution in Chapter 7, Section 3.3, and Part V takes it up where it moves allocation (Chapter 10, Section 3). Other matters bear on the relations as well. Instruments and infrastructure were examined as enabling conditions in Chapter 5, Section 6 and Chapter 7, Section 4.2; Part V takes them up where they realize the three relations, and Chapter 12, Section 6.4 takes up infrastructure as a relation through which effective authority is distributed. Material resources and the formation of researchers are taken up next, together, as support.

Support here names what makes it possible to sustain an organized direction that may fail, such as material resources, the formation of researchers, and the bearing of the costs of failure. It bears chiefly on a further term of the characterization, attempt. At the activity level, obtaining resources, permissions, and access belongs to research when it is undertaken in the attempt to resolve the inquiry (Chapter 7, Section 4.4). Where support ends and research begins thus varies with where the actor's boundary is drawn. Drawn widely, obtaining means and allocating the costs of failure become the actor's own activity, and the actor itself holds the role of supporter. At the level of value and governance, the reasons offered for support are addressed to those who supply it. The supporter and S are distinct roles, which the same party may or may not hold. In a firm's internal research the firm can be both the supporter and the S for which results are new; a patron who maintained a scholar and the community of scholars for which the scholar's results were new need not have coincided; nor, necessarily, do a public funding body and the specialist community or humanity relative to which the research it funds is new. Part V does not count support as a fourth relation. The derivation above does not yield it, since attempt, the term on which support chiefly bears, marks the activity's direction and its possibility of failure rather than where it is tied to S; where attempt enters such a tie, as in the question structure, it does so together with new to S, and the relation that results is allocation; but the derivation does not claim to exhaust the relations around research, so this alone does not settle whether support is to be treated as a relation of its own. Part V does not so treat it because support enters claims about artificial research systems, of the kinds illustrated by the three claims of Section 1, chiefly by way of the three relations. When funding moves which of S's unknowns are pursued, support is part of allocation; when those who judge results also fund the work, support is bundled with certification; when control of funds becomes a power to set and revise the direction of inquiry, support is bundled with authority. Part V therefore takes support up where it realizes or is bundled with the three relations. The question that does not fit these forms, why support is given at all, Section 3 takes up position by position.

2.2 Roles around research

The roles that the three relations involve can be named once. Chapter 8, Section 1 noted that evaluators, beneficiaries, supporters, audiences, and authorities can differ from the actor and from S, and identified them locally where they mattered. Part V uses a few of them throughout, and Table 9.1 states them in one place. Dividing the roles is meant to keep apart parties that the single word "society" can name, such as the reference society, the supporter of research, and the public of a state. The roles are divided by the position in which a party stands to an inquiry: conducting it, being the collectivity for which its results would be new, sustaining it, judging it, relying on its results, bearing its consequences, or making or remaking the actor that conducts it, because these are the positions that claims about allocation, certification, and authority invoke. Several roles can be held by the same party, and in the example of Section 1 an artificial configuration can hold more than one of them.

Table 9.1. Roles around research used in Part V

Role	Who holds it	Where Part V chiefly uses it
A, the actor	the unit to which the conduct of research is attributed: a person, a group, or an artificial or hybrid configuration (Chapter 7)	throughout
S	the society relative to which the intended achievements would be new and for which they are sought; their bearer or mode of realization is a further question (Chapter 1, Sections 2 and 6)	throughout
Supporter	whoever supplies the means of an attempt and bears or distributes the costs of its failure	Sections 2 and 3; Chapter 12, Section 1
Evaluator and addressee	whoever judges, and whoever uses the judgment to choose (Section 4)	Chapters 10 and 11
Those who rely	those who take a certified result as a basis for belief or action, usually but not always members of S	Chapter 11
Those who bear consequences	those on whom the costs and risks of inquiry and of its errors fall, who need not belong to S	Chapters 11 and 12
Provider or maintainer	whoever builds, maintains, updates, or retires the configuration to which conduct is attributed, and so can change the actor itself rather than only its means (Chapter 7, Section 3.1)	Chapter 10, Section 5; Chapter 11, Section 5.4; Chapter 12, Sections 2.2 and 6.4

A further role, the constituency to which a decision about inquiry would be justified, depends on the account of legitimacy adopted and is taken up in Chapter 12, Section 1. S is abstract throughout. It can be humanity, a specialist community, an organization, a mixed group, or a candidate artificial society (Chapter 1, Section 6); Part V does not use "society" for the public of a state unless it says so. The provider is listed apart from the supporter because supplying the means of an attempt and making or remaking the actor are different ways of standing to an inquiry. Section 6 takes up a provider who also supplies the configuration that performs the research as a feature that artificial configurations can bring, one with requirements of its own.

2.3 Realization, division, and recombination

The three relations are analytical relations, not institutions. In the research systems that the book's readers know, familiar institutions have played central roles in them, and those institutions are one realization of the relations among others. Allocation has been carried out through, among other things, investigators' choices of problems under the pressure of credit and careers, the selection of projects by funders, and the direction of laboratories; Chapter 10 examines it. Criticism and certification have been carried out through, among other things, review before publication, replication, the publication of results, and their later use; Chapter 11 examines them, and Section 7 of that chapter takes up the functions of scholarly communication that journals came to bundle and that can be separated and reassigned. Authority has been exercised by, among others, funders, research communities governing themselves, regulators, and publics; Chapter 12 examines it. How each relation is realized also differs by field. In the mathematics group that Laurens Hessels and colleagues studied (Chapter 10, Section 5), arguments could gain recognition before formal publication, and in their humanities group recognition was converted mainly into positions and research time rather than grants; a proof checked against a kernel, the small trusted core of a proof assistant that checks each inference step, which Chapter 11, Section 2 discusses, is a realization of certification that involves no replication; and the authority over a field's standards can lie with its own members more than with funders or regulators. The three relations are meant to cover these realizations as well. The institutions named here are examples drawn mostly from the natural sciences. The view that these institutions are one historically formed realization among others has a lineage. Kitcher, noting that the rules first laid down in the 1660s for choosing investigations and certifying new ideas were adapted to a quite different context, asks whether an institution that had no inkling of its later influence and has "evolved by happenstance" is well designed for the functions it now fulfills (Kitcher 2013, 216). Michael Nielsen and Kanjun Qiu describe the design of science's social processes as contingent on the accidents of history and the space of possible processes as vast and mostly unexplored, and hold that exploring it will take decades or centuries at the least. As reasons why those processes that are collectively held hardly change, they give the centralization of control in a few large funders and institutions, the collective-action problem that no one alone can change a norm, network effects that homogenize social processes through community judgment and the shadow of imagined future judgments, and the absence of a feedback loop that would let new institutions grow. Taking scientists to be users of the discovery system and for the most part not its designers, they envision metascience entrepreneurs who try and spread new processes (Nielsen and Qiu 2022). Chapter 12, Section 9 returns to that exploration, asking what may happen when the configuration that does research can carry it out on its own research. And an essay by this book's author, written with an AI agent, proposed that AI may become material with which the roles and relationships constituting science are taken apart and reassembled differently (Takagi and an AI agent 2026).

What matters here is the structure of relations held by different parties. On the analysis proposed here, the division of these relations among different parties could supply goods that no single party was designed to supply, and Part V examines four of them. The first of the four is error detection across boundaries: points at which an error made in one place could be detected in another. The second is standing for dissent: standing in the sense defined in Section 2.1, held in relation to those who make, or are bound by, a judgment that would settle a question or close a line, to object to it, to refuse to abandon one's own view or line, or to propose that what was closed be reopened, whether or not the dissent later proves right. What the division may have supplied is parties distinct enough to hold such standing and occasions on which it was recognized and exercised; whether an arrangement can supply the entitlement itself depends on the ground of standing adopted (Chapter 12, Sections 5 and 9). What such standing protected epistemically, the continued representation of views and lines that a prevailing judgment disfavors, is a different matter from the standing itself, and the two may come apart; call what it protected the epistemic work of dissent. In criticism, Longino's tempered equality of intellectual authority is one account of what that work calls for (Longino 2025, §5). The third is boundaries across which accounts had to be given, at which one party had to explain or justify what it had done to another. In research, where a result is new to S and there is often no other source against which to compare it, an account of how the result was obtained could be among the few means by which those who rely on it could scrutinize it (Section 4). The fourth is responsible parties: parties who could be held responsible for what was done. Since errors in research are often found only later, this good could include there being someone answerable not only when a result appears but also later, when correction is called for. Separation can also fragment responsibility if no party is answerable for what happens across the boundaries. Chapter 10, Section 5 sets out further functions that human allocation performed without anyone deciding them. The four are meant to give Part V definite goods whose fate can be followed when a relation moves inside a configuration. That the division supplied them is a hypothesis that Part V examines, not a finding, and each good is assigned to the chapters that test it. Error detection across boundaries is examined in Chapter 11, Sections 2 and 3, where two astronomical episodes show it supplied in one case and missing in another; standing for dissent in Chapter 10, Section 4, where dissent takes the form of continuing or reopening a line, and Chapter 11, Section 3, for critics; boundaries across which accounts had to be given in Chapter 11, Section 9 and Chapter 12, Section 6; and responsible parties in Chapter 12, Section 6. Each of these chapters returns at its end to the goods assigned to it, and Chapter 12, Section 9 collects all four.

Restated in the terms of the three relations, the formulation of the essay cited above becomes: an artificial configuration can combine relations held by separate parties. One system can search the literature, generate candidates, choose experiments, evaluate results, route resources, and use the outcome. Combining support, judgment, and direction is not yet the same as bringing the performer inside. A supporter can also judge results and direct work while depending on separate performers. Where those performers can refuse, keep their evidence, or leave, they remain parties whose cooperation is required. A configuration that generates its own investigators can bring the performer inside as well. This is the theme of Part V: relations that in human research were held by separate parties, including the performers of research, can enter one configuration. Recombination can remove delay and loss of information at former boundaries. But it can also remove what the boundaries incidentally supplied. The question of Section 1 can therefore be put more sharply. Of the goods that the division of these relations may have supplied, which depended on premises about human realizers, the people who actually held the relations, with their limited attention and lifespans; which on premises about how those relations have historically been organized, such as careers, journals, and the custody of data, an organization that the view whose lineage was recorded above takes to be reorganizable even with human realizers; which on conditions of inquiry that hold whatever realizes it; and which on relations to parties outside the configuration, including S where S lies outside it? Section 6 states this sorting as the device of Part V.

A relation's entering a configuration is not the same as its being designed and assigned from above. A relation that has entered may be realized by applying, from above, rules that the configuration has set, or it may be realized from below, without anyone setting it, as many machine realizers within the configuration act on their own motives or something equivalent to them. And which realizers it is passed through, and how what passes through them is integrated, can make the same relation realize different things. That design from above becomes possible without passing through human realizers is not the only implication that Part V follows, but it is an important one. In human research, keeping disfavored lines alive, the turnover of participants, the distribution of search by curiosity, and the like have often come about through investigators' own motives and lifespans, without anyone setting them (Chapter 10, Section 5). In a configuration that can be designed from above, by contrast, the likeness and difference among members and the correlation of their errors, as well as what is given to which line, can be set by how it is made (Section 6). What once came about without anyone deciding it may then be set by design, arise from below among machine realizers, or be lost through neither. Whether to adopt curiosity-driven search as a way of allocating may also become harder to leave implicit (Chapter 10, Section 5). In research, moreover, which differences will later prove useful is not known in advance (Chapter 10, Section 2), so that setting diversity, the difference among members and lines, as a value of the design can mean setting it without knowing what it will serve. Nor need that value be set once and for all: the configuration may revise it in light of what its inquiry finds, and the way research is organized then becomes an object of the configuration's own inquiry as well (Chapter 12, Section 9).

Movements in the opposite direction have also been proposed. Sarah Hamburg, writing as an advocate of what Hamburg and others call decentralized science, argues that functions such as peer review and reputation systems "could be governed by separate, specialized communities", which would mitigate "the risk of single platform domination" (Hamburg 2022). Examining decentralization as an ideal, Nathan Schneider argues that such a division can move a center rather than remove it. Surveying uses of the term in political science, in discourse about the Internet, and among advocates of blockchain technology, Schneider finds that decentralization in one part of a system has consistently coincided with centralization in another; in many cases where decentralization is proclaimed, Schneider writes, "the center never departed … it only shifted", and "a theory of decentralization must include a theory of centralization as well" (Schneider 2019, 26, 28). The study of polycentric governance had drawn a related line between form and function. Elinor Ostrom quotes Vincent Ostrom, Charles Tiebout, and Robert Warren, for whom polycentric connotes many centers of decision making "formally independent of each other", whether they actually function independently being "an empirical question in particular cases" (Ostrom 2009, 411). Taken together with the independence powers that Section 5 names, powers such as changing one's own aims, refusing resources, keeping one's own evidence, and leaving with one's records, which let a part diverge from the rest, these arguments suggest that what a further division supplies of the goods above may turn less on how many parties hold roles than on whether the powers that let parties diverge are in fact divided among them. In research the difference can weigh heavily. Since no one knows in advance which lines will succeed (Chapter 10, Section 2), whether lines that a prevailing judgment disfavors survive can turn on whether the parts that pursue them can refuse resources, keep their own evidence, and leave with their records (Chapter 10, Section 4).

2.4 Coupled relations and status for S

Two further points govern the answer to the question of Part V. First, the relations are coupled. Allocation steers in part by signals that evaluation produces (Section 4; Chapter 10, Section 3); criticism changes what is pursued or relied on only where someone acts on it, and who may act on it is a question of authority; and authority decides what allocation and criticism are for. A chapter on each relation therefore also shows where it depends on the others. Second, asking whether a relation can be placed inside a configuration brings in a kind of dependence beyond those Chapter 7, Section 4.2 distinguished. That section separated enabling dependence, through which a capacity exists and is exercised without the external party supplying part of the organization attributed to the actor; directive or substitutive dependence, in which another party supplies part of what is attributed to the actor; and governance constraint. The relations of this part can involve a further kind: dependence that partly constitutes what an actor's outputs count as for S, their status for S. Such a status is not by itself S's having come to know an output; how the two are related differs among the accounts of possession in Chapter 2, Section 6. Whether an output counts as knowledge for S, as a result S may rely on, or as a legitimate direction for S's inquiry may depend on relations outside the producing actor even when nothing outside it enabled, directed, or constrained its production. Whether and when this is so depends on the account of knowledge and of authority adopted, and the following chapters trace the branches.

3. Valuable for S

The value at issue here belongs to the value and governance level of Chapter 1, Section 1: the value that the results of research, and research itself, have for S, that is, which results matter and to whom. The working characterization does not determine which unknowns merit attention (Chapter 1, Section 1), while claims about an artificial research system, such as the claims of Section 1 that it allocates inquiry well, that its certifications may be relied on, or that it may decide for others, presuppose some view of that value; this section sets out the views that such claims can presuppose. This is a different question from the one Chapter 11, Section 4 takes up, whether and how values enter the judgment that a result is supported well enough to be relied on. There, values bear on the acceptance of a result; the distinctions drawn here bear on that question chiefly where it asks whose valuation sets the weight of errors.

The three relations of Section 2 each carry a question of value that the activity level leaves open. Allocation asks which of S's unknowns are worth the effort they would take; certification asks which errors weigh more than others in judging when a result may be relied on; authority asks whose ordering of these matters governs, and who may revise it. This section says what it can mean to call something valuable for S, so that the following chapters can use the distinctions it draws without restating them: Chapter 10, Sections 1 and 6 for the ordering of S's unknowns, Chapter 11, Section 4 for the weighting of errors, and Chapter 12, Sections 1 and 3 for the difference between the right ordering of ends and the standing to impose one.

The phrase "valuable for S" can mean different things in two respects: what the value attaches to, and the sense in which the value is relative to S. In the first respect, the value can attach, for example, to an epistemic achievement, as when understanding is said to be valuable beyond the value of its parts (Chapter 3, Section 7.1); a question, as when some of S's unknowns are said to matter more than others; the form of a result, as when a true result is of little use to members who cannot follow or apply it; the practice of research as a whole, as when one asks what reason S or a supporter has to sustain inquiry at all; or a distribution, as when the same results serve some members of S and not others. Part V is concerned chiefly with questions, the practice, and distribution.

In the second respect, the value can be relative to S in several senses, four of which are these. The first is that it is good for S's members, as on the view, which Mark Schroeder reports Richard Kraut and Christine Korsgaard as holding, that nothing can be good without being good for someone (Schroeder 2025, §1.1.1). The second is that it serves ends S has, as on end-relational theories such as those of Paul Ziff and Stephen Finlay, on which every claim about goodness is relative to ends or purposes (Schroeder 2025, §1.1.3), and as when Kitcher, in Matthew J. Brown's summary of the account, traces practical significance to practical goals (M. J. Brown 2010, 134, 137). The third is that S's members in fact value it. Kitcher's tracing of epistemic significance to broad questions that express natural human curiosity comes close to this sense, although, as Brown reports, Kitcher appeals for that curiosity to our conception of "healthy curiosity" (M. J. Brown 2010, 137–138), which gives the account a normative element as well. The fourth is that a procedure or an idealized deliberation among S's members would choose to pursue it. Kitcher's ideal of well-ordered science is close to this sense for the research agenda: on it, as Longino summarizes it, the decisions actually made about which research to pursue track those that a suitably constituted representative body would make, deliberating with the help of relevant information supplied by experts (Longino 2025, §6; Chapter 12, Section 1). The first two senses follow distinctions that Schroeder's survey of value theory draws, and the third and fourth are taken from Kitcher's accounts of significance and of well-ordered science; the four are set side by side so that the following chapters can say which sense an argument about S's unknowns relies on. This book separates the two respects because an argument about what bears value and an argument about the sense in which that value is relative to S can succeed or fail independently. An argument that one question matters more than another may be convincing, for instance, without settling whether it matters in the sense that answering it is good for S's members or in the sense that a deliberation among them would choose to pursue it; keeping the respects apart prevents an argument about one from being taken for an argument about the other.

When S is a community, the first sense, that it is good for S's members, can be read as concerning the good of its members taken one by one, summed or weighed by a distributive principle, or as concerning a common good, which Waheed Hussain and Margaret Kohn distinguish from aggregate welfare, with or without distributive weighting, by its essential reference to the requirements of a relationship among the members (Hussain and Kohn 2024, §1); the second, analogously, as concerning ends its members hold severally or ends they hold in common. Chapter 10, Section 6 meets the difference when it asks which unknowns are S's to pursue together.

Taken strictly, the fourth sense would make a procedure's or deliberation's choice what makes something valuable. For an actual procedure, the constitutive interpretation of Rousseau's general will that Chapter 12, Section 1 takes up, on which the results of a properly organized democratic procedure are the general will because they emerge from it, can be read as a view of this kind, although Tom Christiano and Sameer Bajaj themselves set it out not as a view of value but among conceptions of democratic authority (Christiano and Bajaj 2024, §3.2.1). As Arnon Keren reports, Kitcher, doubting that an objective conception of well-being can be defended, sought to circumvent the question of what is good by relying on democratic ideals that his society takes for granted, and offered the ideal as a standard against which actual research agendas are measured (Keren 2013, 236–237). The strict reading differs in turn from the view that a fair procedure confers the authority to decide. On that view, a decision reached by the procedure can be binding even if, as a judgment of value, it is mistaken, so the view is a claim not about what is valuable but about who may decide, that is, about legitimacy (Chapter 12, Section 1). That a result is valuable for S in any of the four senses does not make it true, nor is its value evidence that it is true (compare Chapter 2, Section 6). Whether and how S's valuations enter the judgment that a result is supported well enough to be relied on, for instance through the weight given to the errors that relying on it could bring, is the further question left above to Chapter 11, Section 4. The relativity at issue in this section is that of which results matter, and to whom.

Two positions would make the value of knowledge independent of anything particular to S. On the first, a principle that Stephen Grimm attributes to Jonathan Kvanvig and Michael Lynch, every true belief is good for its own sake, at least prima facie (Grimm 2008, 729–730). The claim that every true belief has prima facie value does not by itself rank S's unknowns; it neither assigns them equal value nor supplies a rule for comparing their value. Grimm, who rejects the position, argues that countless truths within easy reach, such as the number of motes of dust on a desk, are not worth learning for their own sake even for an inquirer freed from finitude, and that curiosity attaches not to truth as such but to subjects and to questions of particular types (Grimm 2008, 730–736). On the second position, which Kitcher calls the traditional approach to epistemic significance, the structure of the world fixes which truths are significant, the aim being, for instance, to discover laws, a unified picture of nature, or objective explanations. That position yields an ordering, but not one that comes from S, and which of S's unknowns to pursue would reduce to finding what is significant as one finds anything else. As Brown reports Kitcher's case against this position, once practical and theological answers are set aside it is hard to justify the value of laws, unification, and fundamental causes, and attempts to pick out significant truths through context-independent explanation fail; objective explanation goes on, in Kitcher's words, "only against the background of our questions and our interests" (M. J. Brown 2010, 134–136). On either position, novelty and truth do not by themselves fix which of S's unknowns are worth pursuing, which is one reason the working characterization does not license an instruction to maximize knowledge (Chapter 1, Section 1).

Other positions make the value of knowledge relative to some party's ends, interests, or valuations. Where these are S's, the ordering is relative to S in one or more senses such as the four distinguished above, for example to curiosity and practical projects on Kitcher's account of significance (Chapter 1, Section 8). Which sense is adopted changes what a configuration would have to have access to in order to order S's unknowns: its members' good, the ends S has, what its members in fact value, or what their procedure or idealized deliberation would choose (Chapter 12, Section 1 traces this for the ordering of ends). In research, several of these are not necessarily fixed before inquiry. In Kitcher's procedure, what educates deliberators' preferences is an account of the significance of possible inquiries. Eigi reports that scientific experts provide this tutoring and estimate the prospects of particular projects (Eigi 2016, 67; Chapter 12, Section 4). Where those who conduct inquiry give it, the valuation of S that a configuration consults in ordering S's unknowns may in part be one that the configuration itself has shaped. Value can also be relative to whoever inquires. On Grimm's account, finding out the answers to questions of certain types has a standing value "for us" because such questions are tied to our natural curiosity, which, on this account, is driven by standing prudential or epistemic concerns (Grimm 2008, 733–737). Grimm writes of human inquirers, but applied to a machine A with a curiosity of its own in this sense, the ordering it yields would be relative to A, and where A does not belong to S it need not coincide with any ordering relative to S. A further view cuts across these senses. On this view, which has roots in Dewey's account of valuation, ends are not fixed before inquiry. A judgment of value concerns the consequences of acting on or valuing something, and inquiry into the means of reaching an end, and into the end's own consequences, can lead to a re-appraisal of the end itself (Anderson 2023, §§2.3, 3.1). On the same account, deliberation assigns weights to the things valued in the context of a choice rather than taking them as given, since weights carried over from earlier choices may not fit a new situation, and a value judgment is tested by acting on it and seeing whether its consequences are valued as it predicted (Anderson 2023, §§2.4, 3.2). On that view, which of S's unknowns are worth pursuing is itself open to revision through inquiry, and, by extension, so is the weight given to one kind of error against another (Chapter 11, Section 4); Chapter 10, Section 6 takes up what follows when a configuration changes which questions are tractable.

The practice of research as a whole is also an object of value. On why research is pursued or supported at all there are several positions, five of which are set out below. They come from writings that argued for the support of research, by Vannevar Bush and Abraham Flexner, and from philosophical work on the aims of inquiry, the grounds of scientific freedom, and the presuppositions of empirical knowledge, by Kitcher, Torsten Wilholt, and Max Weber respectively. They are set side by side because they differ in what they ask of allocation. The five overlap, Bush and Flexner each stating more than one, and they are not the only positions. How value enters certification and authority, where the two respects and four senses above serve as well, is taken up in Chapter 11, Section 4 and Chapter 12, Section 1.

On the first, research undertaken without thought of use is valued as the source of uses no one can foresee. Bush held that basic research provides "scientific capital", that the results of any one investigation cannot be predicted although some fraction of such undertakings will prove highly useful, and that discoveries pertinent to medicine have often come "from remote and unexpected sources"; Flexner argued the same from cases (Bush 1945, 9, 13–14; Flexner 1939, 544–548). As Bush drew the consequences, allocation is asked for breadth, developing "the entire front" of medicine and its underlying sciences, and for funds stable enough for long-range programs; because what is valued is the use, Bush also asked for coordinated work on special problems where support could shorten the time before benefits reached the public (9–10, 26).

Flexner's own justification was a second position. Flexner denied that eventual use is the "actual justification" of laboratory work: a new scientific fact, like a poem or a symphony, bears in itself all the justification that institutes of research require (548–550, 552). On this position knowing and understanding are valued for their own sake (compare Chapter 3, Section 7.1). If it is adopted, it can be taken to give a reason to sustain the practice without ordering the lines within it. Like the principle that every true belief is good for its own sake, it does not by itself supply a rule for comparing them, and, since valuing knowledge for its own sake is compatible with valuing it also as a means (Chapter 3, Section 7.1), it does not by itself exclude counting a line's prospective uses; whatever orders the lines comes from elsewhere. Flexner set prospective use aside as a guide to the lines on grounds that go beyond valuing knowledge for its own sake, pleading for "the abolition of the word 'use'", holding that no administrator "can possibly direct the channels" in which investigators work and that the advantages of the discoveries of Pasteur, Koch, and others "could never have accrued if the idea of possible use had permeated their minds" (548–549). In the institute Flexner defended, "each individual disposes of his time and energy as he pleases," and Flexner accepted that some money would be wasted (548, 551).

Kitcher holds a third, developed against what he calls the myth of purity, on which, as Keren glosses it, moral and political concerns are irrelevant to the evaluation of "pure science" (Keren 2013, 240 n. 6). Science aims at significant truths, "correct answers to significant questions", not any truths (Kitcher 2013, 207), and research on apparently recondite questions may earn its significance through the practical gains it prepares, as Thomas Hunt Morgan's concentration on the fruit fly eventually led to the molecular biology that now underlies much biomedical research (Kitcher 2013, 220). Even research whose practical consequences are remote presupposes a judgment of its value relative to other ways of extending inquiry, since far more questions remain than could all be thoroughly investigated, and pursuing some of them would be "a frivolous waste of time and energy" (Kitcher 2013, 219); and, Kitcher adds in a later essay, even where research is relatively invulnerable to moral mistakes, "failure to do the good you can is also a lapse" (Kitcher 2024, 240). If this position is adopted, as Kitcher draws the consequence, no line is exempt from comparison with the practical goals and forms of curiosity it competes with, and the good forgone counts among its costs; the conclusion is not that remote research is unjustified, only that its justification rests on that comparison (Kitcher 2013, 220–221; Kitcher 2024, 240).

On a fourth, research is valued as a source of the knowledge on which citizens form their political preferences. Wilholt reconstructs a political argument for the freedom of research on this ground. On this reconstruction, democratic power is legitimate because its decisions reflect citizens' well-informed preferences, so no political power can legitimately control the generation of the knowledge that feeds those preferences without undermining its own democratic legitimacy (Wilholt 2010, §4). The same reasoning sets the argument's limits: it does not reach research whose results stay out of public discourse; it presupposes a plurality of approaches, since there would be no reason to think that a community that settled on one approach would put citizens in a better epistemic position than one a government favored; it backs a claim to means only in proportion to political relevance, since resources are finite; and, being an argument against political control, it establishes no procedure for setting the agenda (Wilholt 2010, §§5–6). If it is adopted, allocation is asked to keep the agenda independent of those who would govern by the results of research and to sustain plural approaches, and the question it leaves open, who is to define relevance, belongs to authority in the sense of Section 2.1 (Chapter 12, Section 3.2).

Weber, writing on the cultural sciences, held a fifth. Knowledge of an infinite reality by a finite mind rests on the presupposition that only a finite part of it is "worth knowing" (wissenswert), and which part becomes an object of investigation is fixed by value ideas that govern the investigator and the time and change with the culture (Weber 1904; Zeno.org text, 171, 183–184). The belief in the value of scientific truth is itself, on this account, a product of particular cultures, and to anyone who does not value that truth science has nothing to offer (213). Specialized work proceeds without checking its material against those value ideas until the cultural problems that gave it direction move on, and science then changes its standpoint and its concepts (214). If this position is adopted, one can infer that the ordering that allocation follows is not derived from research and is expected to move, so that what is asked is less continual comparison than the capacity to reorient.

What these five positions ask of allocation can change when the investigators are machines. Chapter 10, Section 5 takes this up for Flexner's and Bush's defenses, which can be read as reasons for adopting exploration driven by investigators' curiosity as a way of allocating, whereas whether artificial investigators have curiosity or something equivalent to it cannot be presumed as it could be for human researchers.

In discussions of artificial systems in research, positions of another kind appear, on which research is valued also for what it is to, and does for, those who conduct it. Fulvio Castellacci and colleagues identify what they call a normative fault line: whether scientific institutions exist primarily to produce knowledge or also to develop human capacity, certify expertise, and maintain a system of merit-based allocation. On the output-maximizing view they describe, the effect of artificial assistance on how researchers are trained matters only insofar as learning affects output; on the capacity-development view, substituting artificial systems for human judgment has costs that output metrics do not register. They hold that the choice "cannot be settled on empirical grounds", and the framework they propose adopts the second view (Castellacci et al. 2026, §§2.4, 2.8). In an exchange collected by Marcel Binz and colleagues, Emily Bender, Carl Bergstrom, and Jevin West write that science "is a social process" and that its agents, "real scientists", are "as much the product of this process as the results recorded in papers"; Matthew Botvinick and Samuel Gershman find the prospect of a full-spectrum replacement for human scientists "deeply unappealing" even "in a scenario where technical limitations presented no barrier" (Binz et al. 2025, 6, 8); they also hold that human understanding is to remain a core goal of science (7), and what they would reserve to people is taken up in Chapter 10, Section 6.

John Danaher and Sven Nyholm, asking what automation does to the value of achievement in work, assess an achievement by four variables: the value of the output, whether the agent's causal contribution to it is significant and not lucky, the cost of the agent's commitment, such as effort, time, and attention, and the voluntariness of the agent's action. On their argument, automation can lower the first three together: it reduces the value of the outputs left to human tasks and the cost of human commitment, and it severs the causal connection between human activity and valuable outputs. Of the responses they sketch, emphasizing other aspects of meaningful work, such as mastery and autonomy, may, they note, ring hollow when one's contribution to the output has shrunk, and seeking achievement outside work, in hobbies, sports, or artistic projects, may be limited as automation reaches other domains (Danaher and Nyholm 2021, 230–235). Applied to research, which they touch on only through examples such as Einstein's scientific achievements, the analysis separates two demands: that humans go on taking part in inquiry, with the effort and mastery it involves, which reserving activities for them can meet, and that their taking part remain an achievement, which requires a significant, non-lucky contribution to valuable results and which reserving activities meets only where what is reserved still yields such results. Whether human contributions remain significant to valuable results, however, depends on what others pursue. In Ted Chiang's story described in Chapter 1, Section 6, which is written as a journal editorial, the editorial gives a reason of this kind for continuing human research: human researchers "may discern applications overlooked by metahumans," whose advantages, it says, tend to leave them unaware of human concerns (Chiang 2000, 517). Read as a thought experiment, the case illustrates how the second demand could be met where the society that holds the frontier results does not pursue what matters to the human S, the kind of mismatch Chapter 8, Section 3.2 traces. Yasuo Deguchi calls one threat of replacement a dystopia of the knowledge subject: the loss of the pride humans have taken in being the only intellectual subject. On Deguchi's account, offered as one philosophical option among others, inquiry has always been conducted jointly, so that the solitary or despotic agency supposedly threatened was an illusion, and what would really be lost is the scientist's right of final decision (Deguchi 2026a, 28–29, 31–32). In a statement of the principles of this account, which Deguchi calls the WE-turn, the illusion follows from the thesis that no individual can perform a bodily action alone, a thesis extended there to thinking and deciding; the subject of any such action is then the system of all the agents necessary for it, a system that extends far beyond the body in space and time (Deguchi 2025a, 2–4, 7–9, 12).

These positions differ from the five set out above, from Bush's to Weber's, in what they attach value to. The five value the practice for the knowledge it yields or for what that knowledge does; these value it also as conducted by particular parties. In the terms of Table 9.1 they concern the actor, and the humans who would otherwise conduct research, who are among those who bear the consequences of how inquiry is organized and need not belong to S. Where A is human, the formation of researchers is also among the supports that sustain inquiry (Section 2.1), and an output-maximizing view can value it for that reason. The output-maximizing view and these positions come apart where A can be a machine: the cost these positions register, or at least part of it, is then a change in who conducts research, which can occur even where what S comes to know is unchanged. On these positions, then, two organizations of research that deliver the same knowledge to the same S need not be equal in value, and an evaluation that compares outputs alone cannot settle what is gained or lost when artificial configurations join research. Two further things come apart here: who finally knows, and who is changed in the course of coming to know. The output-maximizing view counts the second only insofar as it bears on future outputs, while a view that values the development of capability can count it in its own right; which is adopted changes how the same replacement is assessed. If one of them is adopted, allocation is asked not only which of S's unknowns to pursue but whose participation in pursuing them to sustain, and, on Deguchi's account, authority is asked whose right of final decision about them to preserve (Chapter 12, Section 2.2).

Several of the positions above, Bush's and Wilholt's among them, were addressed to a state or its citizens. In the terms of Section 2.1 they are reasons addressed to supporters, and the supporter and S are distinct roles, which the same party may or may not hold. Reasons for support raise a question distinct from why the practice is valued: what they require of the arrangement that allocates resources and powers over research. The positions here concern why the practice is valued and what that asks of the ordering of S's unknowns. The two questions meet where a reason for support also constrains its arrangement, as with Bush's report and Wilholt's argument. Chapter 10 asks of each argument about allocation which aim it presupposes.

Justifying support for research and making its results available to S are distinct questions. Partha Dasgupta and Paul David explain why the source of support alone does not settle availability. The knowledge generated in research, they hold, is not inherently available to others: researchers can choose to withhold it from the public domain, and what separates the open pursuit of knowledge from its pursuit under secrecy is the norms of disclosure and the rewards under which research is done, not the source of its funds, so that on their account a publicly funded laboratory that keeps its findings secret belongs with what they call Technology rather than Science (Dasgupta and David 1994, 491–492, 495, 497–498). Inge Kaul and Ronald Mendoza likewise treat nonrival goods, knowledge among them, as goods that "lend themselves to being made public", and count a good as fully public only where all can in fact consume it, as knowledge in written form cannot be consumed by those who cannot read (Kaul and Mendoza 2003, 88–89). These authors write of human researchers and institutions. Applied to a configuration that both supports and conducts research, their accounts suggest that its having obtained a result does not by itself make the result available to the members of S, as in a firm's internal research (Section 2.1): availability requires further conditions, and it turns on what the configuration releases and whether the form of release is one they can use, not on who paid.

4. What evaluating involves

Each of the three relations is realized largely through evaluations: lines are judged worth continuing or not, results are judged good enough to rely on, and the standards and error weights for both are set by decisions that are themselves evaluated against purposes. Part V's question whether relations can be located inside a configuration therefore overlaps largely with the question whether evaluations can move inside, and whether an evaluation can move inside depends on what it consists of, who judges, against which standard, and for whom, and the book has so far used the word without saying any of this. It has used it for claims about artificial researchers (Introduction, Section 3), for the separate judgments of an activity (Chapter 2, Section 8.1), and for appraisal within inquiry (Chapter 5, Section 4). This section therefore sets out the minimal account of evaluation that Part V uses: Chapter 10, Section 3 for the signals by which allocation steers, Chapter 11, Sections 1, 7, and 8 for the evaluation of results and its change over time, and Chapter 12, Section 2 for the decisions that set standards.

The account consists of four points, which do not amount to a theory of evaluation; they say nothing, for example, about how the verdicts of several evaluators are combined. The first two state what the uses above have in common and are not peculiar to research; the last two state features that evaluation takes on in research, and are chosen because the following chapters turn on them.

First, an evaluation judges an object, taken as a thing of some kind, against a standard tied to a purpose, and some party, whether a person, a body, or an artificial configuration, uses the judgment to choose something: to publish, to fund, to rely on, to withdraw. A measurement can be an input to an evaluation without being one, since the evaluation adds the standard. Evaluating also differs from valuing in the sense of Section 3: valuing concerns what is held to matter, and evaluating applies such a valuing to an object for a use. The contrast has a precedent in Dewey's distinction between valuing, or prizing, and evaluation, or appraising, the latter arising when one asks whether something is to be valued (Anderson 2023, §2.1); the use here is narrower. Because standards presuppose purposes, the question "by whose purposes?" leads back to Section 3.

Second, the evaluator and the addressee can differ. In human research, for instance, referees have judged for editors and readers, and panels for funders. A certification is addressed to those who would rely on the result, and an evaluation of support to those who supply it. To whom an evaluation in turn answers is a further question, on which accounts differ, and the roles of Table 9.1 keep apart the parties among whom they locate the answer, or the judgment itself: for instance, on Wilholt's account the judgment of which errors matter answers to the valuations that those who rely on the result expect, whereas Jakob Ortmann holds that judging whether a tool's use is appropriate for a task is a value judgment that can be deferred to a trusted party, which places that judgment with evaluators rather than with those who rely (Chapter 11, Sections 4 and 5.2; Chapter 12, Section 1). Separating what an evaluation does from the arrangements that have bundled several evaluations together also allows Chapter 11, Section 7 to ask what certification for S requires without presupposing the form of review before publication.

Third, evaluation in research relies on proxies, because its target is often not observable when the judgment is made. For instance, later replication and criticism can change warranted attributions of knowledge to S, or whether that knowledge continues to be held; they do not make a previously false claim have been true. Whether a result matters often appears only through later uptake; a prospective evaluation judges results that do not yet exist; and many who rely on a result cannot examine it themselves (Section 1). Judgments, scores, passed checks, and records of later uptake, such as citations in human research, therefore stand in for the target, and, except on accounts that identify the target with such measures, they stand in for it rather than constitute it. That a measure is not the target does not mean that the evaluating it records lies wholly outside the target. On the accounts on which social organization enters warrant or possession (Section 2.1), the criticism, response, and use to which a judgment, a passed check, or a record of uptake belongs can be part of what gives a result the status of knowledge for S or makes it S's knowledge, while the truth of the claim, which such a process does not bring about, remains something a record of it only stands in for. On the account given in this section, the use of proxies follows from the structure of the target rather than from a defect of particular institutions, at least where the target is settled, shows its importance, or comes into being only later; that many who rely on a result cannot examine it themselves depends instead on the capacities of those who rely, and so, for an artificial S, on how evidence is shared among its members (Section 2.1). What matters is how the relation between proxy and target behaves once the proxy has consequences (Chapter 10, Section 3; Chapter 11, Section 8).

Fourth, when what is evaluated is new to S, S cannot simply check a result against what it wanted, since a result new to S answers a question whose answer S did not have. This holds whether the research was commissioned by S or by a supporter distinct from it (Section 2.1; Chapter 8, Section 3.3). Not having the answer is not the same as not having a standard. Where the question fixes in advance what would settle it, as when a formally stated theorem is to be proved under given rules of inference or a registered forecast awaits the event, a new result can be judged against that standard, although whether the standard captures the question remains a judgment of its own (Chapter 3, Section 5.1; Chapter 11, Section 1). Where it does not, the evaluator may not yet have a standard for the result, and a standard learned from past judgments can fail where a new result lies outside their range (Chapter 11, Section 8).

In these terms, allocation is realized largely through prospective evaluation of lines and proposals, and criticism and certification through retrospective evaluation of results, addressed in certification to those who would rely on them; authority is the standing to set and revise the standards and the weights of error by which both kinds of evaluation are made.

5. Control, and the powers that limit it

What is it for a relation to be inside a configuration? In the terms of Chapter 7, a relation is inside when the initiation, monitoring, redirection, and recovery of the decisions that realize it are organized within the configuration's boundary (Section 2 of that chapter), and when both the supply of the work and its authorization and ratification, permitting the work and accepting what it yields, lie there (Section 4.2 of that chapter). A relation can be inside in one of these respects and not the other: in the example of Section 1, the checks may be supplied inside while the authorization to stop or replace the system may lie with its provider. Nor is whether a relation is inside settled by the names or the number of organizations. Two organizations merged into one may leave unchanged who keeps records separately, who can refuse a judgment, and who decides independently how resources are used; conversely, organizations that remain separate may all follow a single judgment. What changes when a relation moves inside is therefore to be asked not of what the boundary is called but of where the organization of decisions, authorization and ratification, and the powers named later in this section that let a part proceed separately come to lie. For authority, which Section 2.1 defined as standing, the decisions that realize it can be organized inside in this sense while whether the configuration has standing to make them remains a further question, which is why the third claim of Section 1 concerns who may decide rather than where the decisions are made (Chapter 12, Section 1). Part V's question thus presupposes a notion of control. One possible answer to it, that humans keep control of whatever moves inside, can be assessed only once it is said what control is. This section says as much as Part V uses: Chapter 10, Section 2 for the philosopher-monarch, Chapter 11, Section 2 for when routes of criticism are independent, and Chapter 12, Sections 2, 4, and 7 for decision rights, standing that is more than nominal, and decisions that remove the capacity to revise.

Control over a relation, in the sense used here, is the capacity to determine what the parties or processes that realize it do. The word can run together four things, which Part V keeps apart: a formal right to decide; the effective capacity to exercise it, which corresponds to control in the sense just defined and which a lack of time, information, or competence, or dependence on another's infrastructure, can empty; knowledge bearing on how the control would best be used, such as the knowledge of each line's prospects that, as Shaw and Viola report, Kitcher gave the philosopher-monarch alongside the prerogative of directing research, which control does not supply and which, on Kitcher's analysis as Viola reports it, still leaves open how the workforce is best divided (Shaw 2018, 155; Viola 2015, 5–6; Chapter 10, Section 2); and influence that shapes the options among which decisions are made without making them, as defaults, rankings, and infrastructure do. The second and third correspond to the questions of effective capacity and epistemic warrant that Chapter 12, Section 2.2 separates. The first is what that section calls formal assignment; whether its holder is entitled to it is that section's further question of authorized power, on which standing in the sense of Section 2.1 bears and whose answer depends on the ground of authority adopted (Chapter 12, Sections 1 and 3). The fourth is added because a party can hold the first three while another shapes the options among which it chooses. The division follows the questions one can ask of a claim that some party keeps control: whether the party holds the right, whether it can exercise it, whether it knows how, and whether something else has already shaped what it chooses among. Its purpose is to let a claim that humans keep control be asked which of the four it asserts, or runs together. The division does not separate, for example, control held at one time from control that persists as a configuration changes, which Chapter 12, Sections 6 and 7 take up.

Against control stand powers that limit it, named once here: the power of a part of a configuration to change its own objective, to refuse resources, to keep its evidence apart, to stop another part's update, and to leave with its record. Each power is included because, if a part holds it, that part's course can diverge from what another part directs; together they give, for allocation and criticism, concrete forms of the authorization and ratification that Chapter 7, Section 4.2 separated from the supply of work. They are the powers that Chapters 10 to 12 use, not a complete list, and because each secures a part's capacity to diverge, Chapter 12 also calls them independence powers. Like control, each power can be held formally, as a right assigned to the part, or effectively, as a capacity the part can exercise, and the two can come apart: a part may be assigned a right to stop another part's update and lack the means to exercise it, or be able to stop the update without any such right. A right can secure a capacity, and a capacity can make a right effective, but neither is the other. Chapter 10, Section 2 asks whether such powers are divided among the hundred instances of Section 1, as a test of whether they are independent of one another, and keeps that question apart from whether each instance, or the whole, conducts research; Chapter 11, Section 2 asks when routes of criticism held apart are independent, which organizational separation alone does not secure; Chapter 12, Section 7 asks which of the powers bear on keeping authority revisable. Rights over particular kinds of decision, including the formal side of several of these powers, are listed once, in Chapter 12, Section 2.2; Chapter 10, Section 2 and Chapter 11, Section 2 turn chiefly on their effective side.

Neither control nor these powers is standing in the sense of Section 2.1. Control and these powers concern what a party has been assigned or is able to determine; standing, what it is entitled to propose, decide, object to, refuse, or revise in relation to others. How the two bear on each other, including whether standing can remain nominal for lack of effective capacity, is taken up in Chapter 12, Sections 1 and 4.

6. How Part V proceeds

Sections 2 to 5 set out the relations, the roles, and what valuing, evaluating, and controlling involve. This section states how Part V uses them: a device for sorting what the division of the relations depended on, a frame for the claims of Section 1, the motive of the thought experiments the part uses, and the order of the chapters.

The device is a sorting, stated here as a grid to be applied rather than as a result; Chapter 5, Section 6.5 has already applied it to the roles of the elements discussed there. Section 2.3 asked of the goods that the division of the relations may have supplied what each depended on. Part V distinguishes four possible answers. Some goods may have depended on premises about human realizers, the people who actually held the relations: for example, that a central director has one person's attention, that allocation passes through investigators' motives, or that turnover comes with human lifespans. Some may have depended on premises about the present organization of science, how the relations have historically been organized: for example, that only the investigator has access to the primary data and the process that produced them, or that review as a gate has rationed scarce editorial attention and publication space and that reviewing has fallen on a few. On the view cited in Section 2.3, such an organization is one historically formed realization among possible ones, and it can be reorganized even while its realizers remain human. Where an artificial configuration removes premises of either kind, the problem they posed may not simply disappear but change form. The two fail differently, however: the first fails when the realizer changes, the second when the organization changes, whether or not the realizer does. Some may have depended on conditions of inquiry, which hold whatever realizes it: for example, on the arguments Chapter 10, Section 2 examines, that no one knows in advance which lines will succeed, or, on the view of criticism that Chapter 11, Section 2 sets out, that independent criticism adds information about errors beyond what existing checks already supply. Chapter 11 distinguishes such an evidential contribution from protection against the loss of that contribution, and from the interpretation of failures (Sections 1–2). Recombination does not remove such conditions, although a configuration can meet them inside itself. And some may have depended on relations to parties outside the configuration, including S where S lies outside it: for example, whose questions stand at the root of what S finds significant, or who may set the terms on which others bear the consequences of inquiry. A configuration can build arrangements under which such relations can hold, but where a relation involves the other party's participation or recognition, it cannot complete it from one side; for relations that do not require the other party's participation, such as what is in another's interest, how far the configuration's side can meet them depends on the account adopted. Below, relations to parties outside the configuration are sometimes called relations to S for short; the short form includes the other parties of Table 9.1.

The four kinds of answer correspond loosely to the elements of a research claim (Chapter 1, Section 2). The first two correspond to who has realized the relations and how, which in human research included parties other than the actor and, once the relations move inside, is the configuration that is the actor: the first concerns the realizers themselves, the second the way their realization has been organized. The third corresponds to the attempt to bring about epistemic achievements, whatever realizes it; and the fourth to S, widened to the other parties of Table 9.1. The sorting is meant to give the question "what changes when a relation moves inside?", for any one good, four forms that its answer can take. If a good depended on premises of the first two kinds, the problem it addressed may disappear with them or remain in another form, and the question is which, and, where it remains, what takes its place; if on conditions of inquiry, a configuration can meet them inside itself, but that it does so needs to be shown; if on relations to parties outside the configuration, what is needed is evidence about those relations rather than an inference from capability alone (see the frame below). One good can, however, fall under more than one kind, and some things that artificial configurations can bring are not premises that fail but features of the new realization with requirements of their own, such as the setting of likeness and difference among members by how they are made (copies of one base system can share errors, while differences can also be built in), a provider who also supplies the configuration that performs the research, or membership that can be multiplied at will (where the multiplied members belong to S, Chapter 8, Section 1.2 noted that an increase in what S knows through research cannot be told apart from one through a change in who its members are unless S's boundary is fixed); the chapters note these where they arise.

That a realization has requirements of its own is not peculiar to artificial ones. Donald Campbell, pursuing a descriptive epistemology, held that knowledge, as such an epistemology studies it, is always embodied in some vehicle whose requirements for keeping itself intact rival the requirement of representing its objects validly, and that a scientific community is such a vehicle: before it can improve its picture of the world it must recruit, reward, and retain members and sustain leadership, requirements that can bias what it carries and that, on Campbell's account, the norms of science, the competitive replication of experiments, and the design of experiments partly counteract (Campbell 1997, 10–12, 27–30). Applied to artificial configurations, this view suggests that removing premises about human realizers, or about how their community has been organized, changes the vehicle rather than removing it, so that a claim that recombination frees inquiry from the biases of human communities can be weighed against the structure and upkeep of the configuration's own vehicle, where features such as those listed above, among them the dependence of members' likeness and difference on how they are made, arise. An artificial configuration, too, is kept up by some arrangement under which it runs, keeps records, and is updated or retired, so the question becomes not only whether the biases of human communities have gone but which conditions of upkeep orient inquiry in their place. Further, where the same evaluation that judges results also decides which investigators are made next and which are retired, that is, where part of the provider's or maintainer's role in Table 9.1 is tied to that evaluation, the ways of asking and the methods that suit the evaluation may be carried over not only into outputs but into the bearers themselves, and the range that future internal criticism can reach may narrow. Campbell's argument alone does not yield this mechanism, but it is suggested for configurations in which evaluation and the generation of investigators are coupled; Chapter 10, Section 3 takes up a related dynamic for allocation that copies the investigators whose measured output is highest. Related forms of the coupling between evaluation and what it evaluates appeared in Chapter 7, Section 4.5, where absorbing an appraisal into a configuration can couple it to the producer's assumptions, and in Chapter 8, Section 3.2, where S's standards move with A; here the coupling reaches the making of the investigators themselves.

Where the lines between the four kinds fall depends on the accounts adopted. Goldman's process reliabilism is an account of justified belief, on which a belief is justified if a reliable process produced it. Applied to certification, as in Chapter 11, Section 5.2, it places much of what certification depends on among the conditions of inquiry, since what would then have to be secured is that the process behind a certified result is in fact reliable, and nothing in that requires a human to run it. On Longino's critical contextual empiricism (Chapter 11, Section 3), certification also depends on relations within and between the relevant critical communities. On this account evidential relations are mediated by background assumptions, and critical interaction within and between scientific communities is the only check against the arbitrary dominance of subjective preference in accepting such assumptions; that interaction must meet norms that apply to communities, namely venues for criticism, uptake of criticism, public standards, and tempered equality of intellectual authority; and what counts as knowledge is determined by such interactions (Longino 2025, §5); what certification has to secure then depends on the perspectives available to the accepting community and on how its criticism is organized. Where that community is S, these are relations to S; where it lies inside the configuration, whether and how S's perspectives enter its criticism is a further question (Chapter 11, Section 3). If significance were fixed by the world, which unknowns to pursue would belong with the conditions of inquiry; on Kitcher's account of significance, by contrast, it would belong with relations to S. Each of the following chapters applies the sorting to its relation and records where the lines move, and Chapter 12, Section 9 collects the results.

The frame concerns the three claims of Section 1. A claim that a configuration allocates inquiry well, that it certifies its results for S, or that it may direct inquiry on behalf of others is a claim about the relations of Section 2. To the extent that those relations fall, on the account adopted, among the relations to parties outside the configuration, such a claim needs evidence about those relations rather than an inference from capability alone. For allocation, the question is whether the signals it steers by still trace back to S's questions; for certification, whether its certifications withstand criticism that adds relevant information and whose findings remain open to scrutiny, and, on some accounts, whether they are acceptable relative to S, whether those who rely can obtain reasons for relying, and whether someone answers for the result; for authority, whether it stands to those who bear the consequences of its decisions in the relation that the ground of authority adopted requires. On democratic authorization through actual participation, the authority exercised rests on the participation of those affected, and on republican accounts on the democratic reading of non-arbitrariness, power is non-arbitrary only to the extent that those subject to it control it; on these grounds, the latter by an inference from the accounts as Chapter 12 surveys them, their standing to object, refuse, or revise is to be preserved as such. Kitcher's standard judges an arrangement by its outcomes, and the fallibilist form of the result-based ground counts the capacity to detect and correct institutional error as part of what makes results good; on these grounds standing matters as a means, for instance as a channel through which members' viewpoints and failures can become known. On the simple form of the result-based ground, which asks only whether results are better, it is not required, and what is needed is evaluation able to show whether the results are in fact better (Chapter 12, Sections 3 and 7). Whether and how far each of these holds is what Chapters 10 to 12 examine, and each ends by saying what its examination implies for the corresponding claim. Which S is specified also sets the scope of a claim. If a firm is S and the claim concerns knowledge new to the firm, more of the relations can lie inside the firm's own arrangements, although those who bear the consequences may still lie outside it; a claim of knowledge new to humanity widens the parties whose relations are in question. If S is an artificial society made largely of the configuration's own instances, most of S's members lie inside the configuration, so that much of what relations with them require can be met there and falls outside the fourth kind, while those who bear the consequences of its research may still lie outside it; Chapter 12, Section 5 takes up that case, including the limit at which no one outside it would bear them.

Part V also uses four thought experiments: Kitcher's philosopher-monarch, who has both complete control of the scientific workforce and an eye for the objective merits of theories (Chapter 10, Section 2); a republic of instances, many investigators derived from one base system (Chapter 11, Section 3); a society that stands as a layperson toward the configurations that hold its evidence (Chapter 11, Section 6); and the limit of a single configuration that is both the actor and the whole of S (Chapter 12, Section 5). Several of these draw on philosophers' idealizations of human inquiry, devices for isolating what a premise about human realizers, or about how their inquiry has been organized, contributes. As artificial systems come to conduct research, the premises of science as it has been organized may change, and ideas that were once only thought experiments can return as possibilities worth examining, or as vantage points from which to think about how research might next be organized. The essay by this book's author cited in Section 2.3 stated this motive earlier (Takagi and an AI agent 2026). Chapter 8, Section 3.4 gave one precedent for the form such a return can take. Feyerabend, holding that incommensurable theories may have no comparable consequences, proposed to judge a theory's predictive success by whether a robot without sense organs, with the theory built into it, reacts in a domain as beings trained to find their way about it do, a criterion that makes no reference to meanings and so bypasses translation between theories (Feyerabend 1962, 94–95). A related route has recently been tried in the bounded domain of chess, with concepts extracted from an artificial system: in a small study that its authors present as a proof of concept, four grandmasters solved more unseen puzzles after studying positions that exhibited those concepts, without labels or verbal descriptions (Schut et al. 2025, §§1.2, 3.2, 4.1).

Otto Neurath gave an argument, for economic orders, for examining arrangements that exist only in thought. Neurath proposed calling utopias all orders of life that exist only in thought and image, without the word's saying anything about whether they are possible, and treating them as constructions of social engineers that belong to theory itself, to be developed and examined side by side rather than one at a time; Neurath also held that "one fully recognizes the real only when one also surveys the possible" (Neurath 1919, 228–231), which can be read as a reason for examining them together. Read with research in view, some of Neurath's premises appear specific to the economy: that the quantities of production and consumption can be calculated in kind and regulated, and that social engineering, like pure mechanics, deals with well-known elements and meets no new forces (228–229, 231). Neither can be assumed for research: on the arguments Chapter 10, Section 2 examines, what inquiry will yield is not known in advance, and artificial configurations can bring features with requirements of their own, of the kind noted above. Others do not depend on the economy, three of which are these. One is that the present order is accidental, so that a theory confined to it and to the past cannot cope with upheavals such as those of war and revolution (231), a form of the view whose lineage Section 2.3 recorded. Another is that for a given aim several equally correct solutions can be found under definite presuppositions (229). The third is that a construction must take account of the properties of those who are to realize it, as an engineer takes account of the elasticity of iron and the breaking strength of copper (229). Neurath did not consider these three for the organization of research, and they are not claims of the same strength. This book takes the first and the third to hold for it as well, the third because, as Campbell's argument above suggests, whatever realizes an organization of inquiry, human or artificial, brings properties that the organization has to take into account. The second, that several equally correct solutions can exist, holds or fails depending on the aim and the standard of comparison, and the book treats it as a possibility to be examined for the organization of research. On these premises Part V uses those of its thought experiments that remove a premise about human realizers: a configuration without such a premise is examined in order to see what the existing division of the relations supplied. Two further elements of Neurath's view are not taken over, or are taken over only in part. One is the wish that such constructions guide action, as party programs and as prophecies that help to bring themselves about (228–229), which is not taken over. The other is the premise that a large part of the order of life can be shaped toward a goal (228), which holds here only in part, since whether what was supplied without anyone deciding it can be set by design is what Chapter 10, Section 5 asks, and since those relations to parties outside a configuration that involve the other party's participation or recognition cannot be completed from one side. Each thought experiment in Part V states where it appears which premise it isolates: the monarch separates control from the eye for the merits of theories; the republic separates the number of critics from their independence; the society as layperson separates deference on matters of content from standing over ends; and the limit separates what depends on relations to parties outside the configuration from the conditions of inquiry.

The chapters follow the three relations of Section 2: Chapter 10 takes up allocation, Chapter 11 criticism and certification, and Chapter 12 authority, ending with what Part V has found; each opens by saying how its sections proceed. The order has two reasons. One is that allocation and certification are first-order relations and authority second-order (Section 2.1). The other is that, in the terms of Section 4, allocation works largely through prospective evaluation and certification through retrospective evaluation, while authority is the standing to set the standards of both, so that it is taken up after them. Of what Chapter 7, Section 5 left to this part, power is taken up in Chapter 12, Sections 2 and 6; oversight in Chapter 12, Section 4; responsibility in Chapter 12, Section 6; and correction in Chapter 11, Section 8 and Chapter 12, Sections 2 and 6.

The hundred-instance example now poses three connected construction questions. Can its allocation preserve productive alternatives while revising what it pursues? Can its criticism give reason to rely on its results, and to whom? Who may authorize and revise its direction, including the creation of further investigators? Some things already come into view from this chapter's preparation. One is that the quality of a research system cannot be read off the quality of its individual researchers: applied to artificial systems, the results of Mayo-Wilson, Zollman, and Danks imply that an agent's performance assessed alone does not show its performance among other agents with whom it shares evidence (Section 1). Another is that even where an artificial configuration delivers the same knowledge to the same S, the organization of research need not keep the same value: on positions that value the formation of researchers, who took part in the process remains in question; applying Danaher and Nyholm's analysis of achievement, whether human participation remains an achievement; and on Deguchi's account, who holds the right of final decision (Section 3). A third is that, on some accounts, what an actor's output counts as for S can depend on relations outside it even when nothing outside it enables, directs, or constrains its production (Section 2.4). A fourth is that, applying Campbell's view, removing premises about human realizers or about how their communities have been organized changes the vehicle that carries knowledge rather than removing it, so that the question becomes which conditions of upkeep orient inquiry in their place (Section 6).

Answers to the three questions are sought also because they make it possible to describe what may happen to research when the relations move inside an artificial configuration, and this chapter's preparation already yields an outline of changes that the following chapters test. First, in the terms of Table 9.1, S may move from the position of evaluator to that of addressee. A result new to S cannot simply be checked against what S wanted, and often there is no other source against which to compare it (Section 4; Chapter 12, Section 4); where criticism and certification both lie inside the configuration and S's members stand as laypeople toward it in every field, what falls to them may chiefly be to use the configuration's judgments in choosing whether to rely on its results (Chapter 11, Sections 6 and 9.3). Second, the direction of research may drift toward what the configuration can readily evaluate. Allocation steers by signals that evaluation produces (Section 2.4), and evaluation in research relies on proxies (Section 4), so success on tasks that are easy to judge may justify the next allocation and narrow from within the questions pursued (Chapter 10, Sections 4 and 6). Third, diversity may become a variable of design: differences among lines and views that in human research arose without anyone deciding them may be set by how the configuration is made, or lost, and this happens without knowing which differences will later prove useful (Section 2.3; Chapter 10, Section 5). Fourth, where evaluation is coupled with the generation of investigators, the ways of asking that suit the evaluation may be carried into the bearers themselves, narrowing the range that future criticism can reach (this section; Chapter 11, Section 9.3). Further, what these four prospects turn on, such as who criticizes, by what standard, and how lines are assigned, need not stay where someone first set it. If an artificial configuration generates critics, relearns standards, and revises how it assigns lines as its inquiry proceeds, the way research is done changes within that inquiry, and the possibility that the evaluated side produces the reference points of its own evaluation, that against which it is evaluated (critics, standards, the society appealed to, and the like), may then appear as a form running through all four; and on several of the accounts examined in the following chapters, norms stated for communities whose critics, alternatives, and peers the party judged met rather than made could then be met by what the configuration itself makes (Chapter 12, Section 9). Chapters 10 to 12 take up the three questions above under the accounts set out here. Their comparisons will also test what separate parties in human research supplied, and which of those functions artificial organization might preserve, change, or lose.