Epistemic Policy — Best Explanation and Open-Ended Inquiry

Working translation v0.2 · 2026-09-06

1. Scope — from “what counts as knowledge?” to “how should inquiry proceed?”

This page does not attempt to define necessary and sufficient conditions for knowledge, justification, or truth from scratch. Its narrower concern is how a finite inquirer should provisionally adopt a world model, criticize and revise it, and remain open to new possibilities when evidence, concepts, and represented alternatives are incomplete. Its center of gravity is therefore an operating policy for inquiry rather than a taxonomy of static belief states.

In this respect, the page shares the orientation of contemporary zetetic epistemology: epistemic evaluation need not stop at belief states at a time, but may extend over the temporal process of inquiry—question formation, evidence gathering, generation of alternatives, assessment, provisional settlement, suspension, monitoring, and reopening. A belief may be rationally held while permanent closure of the corresponding inquiry remains unwarranted.

The project does not, however, build “good inquiry,” truth-seeking, or openness into an ultimate normative foundation. These can first be treated as useful epistemic policies for agents that presently aim at truth discrimination, error correction, prediction, intervention, and self-correction. Whether inquiry itself has independent objective normative value is a separate question. The normative status of inquiry norms remains part of value inquiry rather than an exempt premise.

2. Starting point — epistemic asymmetry without ontological privilege

The Momentary Epistemic Minimum does not place the diachronic self, the external world, the brain, or physical law in the minimal epistemic layer. At most, it identifies the very thin limit that some present epistemic occurrence or constraint cannot simply be eliminated. This asymmetry does not preserve mind as a second substance.

What is special here is not a metaphysically privileged entity called “mind,” but an asymmetry of access: present cognition is the entry point from which a world model is formed. Memory, personality, diachronic selfhood, and even categories such as “mental” or “phenomenal” need not be included in the minimum. They are already higher-level hypotheses about how the present occurrence should be classified and causally situated.

Within the current best world model, the epistemic occurrence itself is best treated as part of a physical process involving brain, body, and environment. The present position can therefore be described as provisional physical monism at the ontological level, combined with an epistemic asymmetry of present access. But this monism is not deduced from the epistemic policy itself; it is the present output of applying the policy to current evidence.

3. Inference to the best explanation is the default where explanatory comparison applies

Above the epistemic minimum, where explanatory comparison is appropriate, the project uses Inference to the Best Explanation (IBE) as its default policy of model selection. This does not reduce logic, mathematics, conceptual analysis, or the epistemology of normativity to a single explanatory score. In particular, IBE does not by itself determine what should count as evidence for normativity.

Competing models are assessed by criteria including fit with observation and experience, predictive power and novel predictive success, successful intervention, explanatory scope and unification, simplicity and resistance to ad hoc repair, coherence with other well-supported knowledge, and robustness across independent measurements, methods, and models.

These criteria need not collapse into a fixed scalar “explanation score.” Simplicity may conflict with data fit; broad unification may trade off against local precision; conceptual economy may conflict with the need for genuinely new categories. IBE is therefore better understood as a fallible comparative policy for integrating several forms of epistemic success than as a finished function of the form argmax ExplanationScore(H).

Constructive empiricists such as van Fraassen press a further question: even if a theory explains well, why assume that explanatory virtues are truth-conducive with respect to unobservable structure? This project does not ground IBE as an a priori truth-tracking principle. It adopts IBE provisionally because explanatory, predictive, interventionist, and integrative success make it a powerful current method. IBE itself is therefore not exempt from the policy of revision.

“Epistemic commitment” here primarily means high credence together with acceptance of a model as the working default. It does not by itself imply irreversible practical action or institutional lock-in. When evidential differences are large, very high confidence in one model is appropriate. Fallibilism is not neutrality.

4. Coherence is not enough — Haack, theory-ladenness, and external constraint

Susan Haack’s foundherentism compares evidential support to a crossword puzzle. An answer must fit intersecting answers, but it must also fit the clue. For this project, the analogy captures the joint role of internal coherence and external constraint.

A perfectly coherent world model is insufficient if it is insulated from experience, measurement, prediction, and intervention. Yet individual sensory inputs need not be treated as self-interpreting, infallible foundations either. External constraint and the web of belief mutually calibrate one another.

“External constraint” does not mean that theory-independent raw data arrive transparently. Scientific observation is mediated by instruments, calibration, statistical processing, classificatory schemes, and background theories. Theory-ladenness does not erase the resistance of the world, but it does mean that the interpretation of what was measured, what counts as an anomaly, and which assumption failed may itself be revisable.

This is also where the Duhem–Quine problem matters. Predictions usually follow not from a central hypothesis H alone but from H together with auxiliary assumptions A, measurement assumptions M, and initial conditions I. If H + A + M + I → E fails, H alone is not mechanically singled out as false. External constraint is therefore best understood as persistent resistance across multiple measurements, auxiliary assumptions, replications, and interventions rather than as a simplistic one-observation falsification rule.

5. Best explanations must remain exposed to criticism and error detection

Explanatory attractiveness alone can favor elegant stories or theories flexible enough to accommodate anything. IBE is therefore supplemented by Popper/Bartley-style criticizability, severe testing, replication, and robustness analysis.

Falsification does not automate theory choice either. As the previous section notes, auxiliary hypotheses, measurement error, approximation, and statistical noise intervene in actual tests. A single failed prediction does not uniquely determine what should be abandoned. A Lakatosian perspective adds a temporal criterion: is a research programme continuing to generate novel predictions, explanations, and problem solutions, or is it increasingly protected only by retrospective repairs?

From Deborah Mayo’s severe-testing approach, the project can borrow the idea that merely “passing a test” is weak evidence if an importantly false hypothesis would also have passed easily. A stronger test is one with substantial capacity to expose the relevant error. The point here is not wholesale commitment to one statistical philosophy, but a separation between explanatory attractiveness and error-detection capacity.

The division of labor is useful: IBE primarily addresses what to provisionally adopt; critical rationalism helps address how to keep adopted views corrigible. From Bartley’s pancritical rationalism, the project especially borrows the meta-principle that epistemic methods themselves are not exempt from criticism. Criticizability alone, however, does not provide a sufficient positive ranking among competing theories, so explanatory comparison and empirical testing remain necessary.

6. Present acceptance and future correction can coexist — what to borrow from Isaac Levi

Isaac Levi’s pragmatist epistemology shifts attention away from repeatedly justifying the pedigree of current beliefs and toward the rational revision of a corpus of belief. It also distinguishes strong current acceptance from the possibility of later correction.

The project does not adopt Levi’s infallibilism. It readily allows an agent to strongly believe a hypothesis while assigning positive credence to represented rivals. But the structural distinction between current acceptance and future corrigibility is important.

For this project, at least three layers should be separated. First, graded credence in proportion to evidence. Second, working acceptance, in which one model is used as the normal premise for calculation, prediction, and research. Third, practical commitment, in which resources, institutions, or future capabilities are changed in ways that may be difficult or impossible to reverse. For example, using H as the working model at P(H)=0.97 does not imply making future consideration of not-H impossible.

Levi’s notion of serious possibility is also useful, provided that represented serious alternatives are distinguished from unconceived alternatives that are not yet part of the agent’s representational space. This distinction becomes important in separating uncertainty to which ordinary credence can be assigned from uncertainty not yet represented inside the current probability space.

7. The hypothesis space is not closed — pessimistic induction, unconceived alternatives, and Bayes

The pessimistic induction from the history of science does not by itself prove that present successful theories are false. Current theories may rest on richer evidence, more precise measurement, and stronger theoretical integration than their predecessors. But the history still cautions against turning present success directly into a claim of finality.

P. Kyle Stanford’s problem of unconceived alternatives is more directly relevant. Defeating all known competitors is not the same as defeating alternatives that have not yet been conceived. Epistemic states should therefore distinguish at least (A) the currently accepted theory, (B) currently represented rivals, and (C) alternatives not yet represented at all.

Category C is not always just a low-probability hypothesis. If the content of a hypothesis has not yet been represented, its credence and likelihood may be undefined in the current probability space. Recognizing unknown possibilities is not the same as assigning arbitrary numerical probabilities to them.

Bayesian Updating and Open Hypothesis Spaces develops this point in detail. If the current candidates are H1,…,Hn, one can preserve probability mass outside them with a catch-all H* = ¬(H1∨…∨Hn). But then the question remains how to determine P(E|H*), and how to represent the internal structure of what is still unknown. A catch-all can be a container that prevents the unknown from collapsing to zero without constituting a full model of the unknown.

The project therefore treats Bayesianism in two layers rather than rejecting it. Within represented hypothesis spaces, Bayesian updating remains a central local updating method; for the hypothesis space itself, the project separately preserves open-world generation, repartition, and reopening. Rational updating inside a set of candidates is not the same achievement as demonstrating that the candidate set exhausts reality.

8. Concepts, measurements, and models are iteratively revised — Hasok Chang

Hasok Chang’s epistemic iteration describes scientific progress that begins from imperfect measurement practices and conceptual systems and then improves those starting points by using the results they make possible. Science need not wait for an independently certified foundation before inquiry begins.

A bootstrap problem appears here. Reliable theories seem to require trustworthy measurement, while trustworthy measurement depends on theories and concepts that tell us what to measure and how. If inquiry required completely independent certification of every starting point, it could never begin. Actual science can instead use a provisional bundle of measurement, concepts, and theory, then recalibrate that bundle through its successes and failures.

This fits the project’s relation among world models, conceptual schemes, models of agency, and models of value. Inquiry must begin with concepts already available, but those concepts may be rewritten by the inquiry they enable. World model ↔ conceptual scheme ↔ inquiry method is therefore a mutually revising relation rather than a one-way pipeline.

Using a present category is distinct from treating it as a final joint in nature. In value inquiry as well, current vocabulary such as “desire,” “reason,” “subject,” or “experience” may itself be reconstructed by later inquiry.

9. Pluralism is an error-correction and exploration strategy, not relativism

Maintaining multiple models, research programmes, measurement strategies, and research communities can have epistemic value. This does not require assigning equal credence to them.

Lakatosian research programmes, Longino-style critical interaction, and convergence across independent methods reduce different forms of correlated error. If methods resting on different auxiliary assumptions converge on the same conclusion, the chance that the result is merely an artifact of one model is reduced. Dissenting research paths may also expose blind spots shared by the dominant approach.

Pluralism also has costs: dispersion of resources, long-term maintenance of low-quality hypotheses, delayed consensus, and the opportunity cost of sustaining alternatives after the balance of evidence has shifted strongly against them. The project therefore does not claim that all alternatives must be preserved forever. What should be maintained is epistemically productive diversity whose error-correction and alternative-generation benefits are reasonable relative to residual uncertainty and maintenance cost. The concrete allocation problem belongs to decision theory.

10. Applying the policy to present ontology — objects and provisional physical monism

The project does not place categories such as “object,” “subject,” “process,” or “relation” in the epistemic minimum. Ontological commitment to an entity is strengthened when models containing that entity succeed in explanation, prediction, intervention, and independent testing.

Objects are therefore individuated within world models, but this does not imply idealism according to which models make them exist. Hurricanes, biological organisms, and corporations, for example, need not be primitive entities of fundamental physics in order to function as stable units of prediction, tracking, and causal explanation. Treating something as real is therefore distinct from claiming that it is metaphysically fundamental.

The project currently treats mind-independent reality and one causally integrated physical world as the best explanation of the available evidence. Mind, cognition, and consciousness are presently modeled as processes within that world. But this is not an axiom of the epistemic policy. It is a current output of epistemic policy + current evidence → provisional physical monism.

The policy can therefore survive a future revision of physicalism if a better explanation emerges. It likewise leaves open whether the deepest structure of reality is best described in terms of objects, processes, relations, structures, or categories not yet available to us.

11. Belief rationality and inquiry rationality are distinct — the zetetic turn

Very high confidence in a hypothesis H is not the same claim as saying that inquiry into H should be permanently terminated. As the zetetic turn emphasizes, the rationality of belief and the rationality of inquiry can be distinct objects of evaluation.

“Keeping inquiry open” does not mean actively rethinking every question at all times. Once evidence is sufficient, a question may be treated as operationally settled so that calculation and action can proceed. The important distinction is between such provisional settlement and terminal closure that makes reconsideration impossible.

Inquiry states are therefore better represented by something like active inquiry → provisional settlement → dormant monitoring → reopening than by a binary continue/stop distinction. New evidence, persistent anomalies, a strong rival theory, a new measurement capability, or serious criticism of background concepts can function as reopening triggers. This avoids the cost of repeatedly restarting settled debates while preserving corrigibility.

For example, using physicalism as an almost certain working world model today is compatible with preserving the capacity to consider a radically different explanation tomorrow. The project does not infer an unconditional duty to continue every inquiry forever. Costs of inquiry, competing aims, stopping conditions, and the value of preserving reopening capacity belong partly to decision theory.

12. The epistemic policy itself is revisable

IBE, Bayesianism, critical rationalism, present statistical methods, present scientific methods, and even concepts such as “explanation,” “evidence,” and “object” are not exceptions to the policy.

This rule is not itself exempt as a self-justifying final principle. If “all methods must remain revisable” were treated as the one infallible foundation, it would reproduce the very exemption it rejects. The project instead adopts this meta-policy provisionally in light of the history of inquiry, the role of error correction, and the usefulness of self-revision.

Nor does this require an infinite hierarchy in which every revision rule must first be justified by a higher revision rule. In a Bartleyan spirit, the key demand is criticizability rather than ultimate justification. A current rule R can be used while remaining open to transition through Rₜ → criticism / new evidence → Rₜ₊₁.

This is not methodological skepticism. The best-performing methods available now should be used strongly now. The point is only that future cognitive expansion, new forms of evidence, or new inferential capacities may reveal better epistemic methods. Strong present use and permanent methodological fixation are different things.

13. Boundary with decision theory

This page primarily concerns what to believe, how strongly to believe it, what to accept as a working model, and what to keep criticizable or revisable. Questions such as how much to spend on further inquiry, whether to delay irreversible action, how to value information and option value, or how to act under Knightian uncertainty and imprecise probabilities belong to a separate layer of decision theory.

Value should not be placed downstream of decision policy. Inquiry about the world and inquiry about value are better treated as distinct inputs into decision-making.

Action and experiment then generate new evidence and feed back into inquiry, so the structure is a loop rather than a one-way staircase. Value inquiry is both a special domain of inquiry and an input required for decision-making. Just as strong practical commitment to present values need not imply permanent value lock-in, strong epistemic commitment to the present world model need not imply permanent epistemic lock-in.

14. What this page does not claim

• That IBE applies in the same form to every epistemic domain, or that it is an a priori or infallible truth-tracking principle.
• That present scientific realism or physicalism is finally established.
• That the mere possibility of unconceived alternatives should greatly reduce confidence in well-supported present theories.
• That every alternative research programme must be preserved indefinitely.
• That Bayesianism is invalid; it remains a central tool within represented spaces.
• That inquiry, truth, or epistemic openness is itself an ultimate objective good.
• That current categories such as “mind,” “object,” “subject,” or “process” capture the final joints of reality.

Selected references and entry points

• Susan Haack, Evidence and Inquiry (1993) — foundherentism and the crossword analogy.
• Karl Popper, Conjectures and Refutations; William W. Bartley III, The Retreat to Commitment — criticizability and pancritical rationalism.
• Isaac Levi, The Enterprise of Knowledge (1980); Pragmatism and Inquiry (2012) — belief corpus, serious possibility, and corrigibility.
• Hasok Chang, Inventing Temperature, ch. 5 (2004) — epistemic iteration.
• P. Kyle Stanford, Exceeding Our Grasp (2006) — unconceived alternatives.
• Sylvia Wenmackers & Jan-Willem Romeijn, “New Theory about Old Evidence: A Framework for Open-Minded Bayesianism” — catch-alls and open-minded Bayesianism.
• Bas C. van Fraassen, The Scientific Image (1980) — constructive empiricism and pressure on IBE.
• Deborah G. Mayo, Statistical Inference as Severe Testing (2018) — severe testing and error detection.
• Benoit Guilielmo, “Inquiry, Questions, and Actions” (2024) — inquiry dynamics in the wake of Friedman and the zetetic turn.
• Helen Longino, Science as Social Knowledge (1990) — social criticism and objectivity.

Related pages

• The Momentary Epistemic Minimum
• Can Pragmatism Ground Value Inquiry?
• Can Coherence Support Normative Knowledge?
• Bayesian Updating and Open Hypothesis Spaces
• The Closed-World Problem in Moral Uncertainty
• Allocating Inquiry Under Unresolved Normative Uncertainty
• Reflective Uncertainty and Irreversible Commitment