Can Coherence Support Normative Knowledge? — Coherentism, Reflective Equilibrium, and Scientific Epistemology
1. First reconstruct reflective equilibrium fairly
Reflective Equilibrium (RE) is not identical to coherentism. RE is a method of mutual revision among judgments, principles, and background theories. Coherentism is an epistemological view that places justificatory weight on mutual support within a web of belief. Using RE does not by itself commit us to the claim that coherence alone generates knowledge.
Goodman's prototype mutually adjusts particular judgments about correct inference and rules of inference. Rawls extends the method to justice by moving between considered judgments and principles. In wide reflective equilibrium, associated with Daniels and others, background theories, rival theories, and relevant empirical knowledge enter the revision process as well.
So it is inaccurate to describe RE as a vote among present intuitions. General principles can defeat particular judgments; resilient judgments can force revision of principles; background theories themselves can be replaced. RE is a powerful form of internal error correction.
It is also important that Consistency is not Coherence. Coherence can include mutual support, explanatory integration, comprehensiveness, simplicity, and robustness under counterexample. The real question is what remains unestablished even after coherence in this thicker sense has been achieved.
2. The central issue — what does coherence actually achieve?
At least five levels should be distinguished:
- Consistency: few explicit contradictions.
- Coherence: judgments, principles, and background theories support one another within an integrated model.
- Understanding: the agent sees its commitments, presuppositions, and implications more clearly.
- Justification: the agent has epistemic entitlement to retain the belief.
- Truth / knowledge: the agent actually represents normative facts or reason-structures correctly.
Equilibrium → Understanding is highly plausible. Equilibrium → Truth is a different claim. The central provisional conclusion of this page is that RE is an excellent optimizer / understanding-generator, but not by itself a verifier.
If this distinction is blurred, an internal achievement — being maximally reflective and coherent — can slide into an external claim — being normatively correct. The deeper problem for coherentist approaches is therefore not circularity as such, but why the circle should be expected to track truth.
3. Why can error be coherent?
RE needs inputs. Rawls filters considered judgments by conditions such as calm reflection, information, and reduced pressure from immediate interests. But once we write J → Filter → J*, a prior epistemological question appears: why trust the filter? If the input judgments have no epistemic standing at all, coherence alone need not create such standing.
Judgments can also have genealogies of the form Evolution + Culture + Learning + Emotion + Institution + Power → J. Such a genealogy does not show that J is false. But causal origin is not truth tracking. If our having J is well explained by evolution, socialization, coordination, or institutional incentives without appeal to normative truth, then using J as evidence requires further argument.
The same point appears as multiple equilibria. Different starting points may yield RE(J0) → E1 and RE(J'0) → E2, each highly coherent. If education, institutions, and power shape the inputs, we can also get self-reinforcing dynamics such as Institutiont → Jt → REt → Principlest → Institutiont+1. High coherence may therefore reflect social endogeneity rather than truth tracking.
Convergence is likewise underdetermined. Convergence does not entail truth. When agents converge, we must distinguish constraint by an independent normative reality from convergence generated by shared psychology, culture, coordination incentives, or power.
As an extreme null model, imagine a world with no objective normative reasons at all. Consistency, coherence, reflective revision, equilibrium, convergence, and moral understanding could still exist. So coherence by itself is not evidence that normative reality exists.
4. Street — can coherence constitute reasons?
Sharon Street offers one of the most important responses to this problem. On her Humean constructivist picture, correct judgments about reasons do not have to track independently existing normative facts. They are constituted through what withstands scrutiny from within an agent's evaluative standpoint — the agent's actual set of evaluative and reason-involving judgments. This starting point is thicker than a mere list of desires.
The attraction of this move is clear. If normative truth is not a practice-independent object but instead consists in what follows correctly from an agent's evaluative standpoint, then RE need not be supplemented by an external moral sensor. Street responds to genealogical pressure not by searching for evolution-free intuitions, but by offering a constructivist account of reasons.
That response, however, puts pressure on what “coherence” is doing.
First reading — coherence as instrumental or constitutive constraint. If an agent decisively values X and understands that Y is necessary for X, refusing to acknowledge Y as a reason is inconsistent with the agent's own evaluative standpoint. This is a powerful internal constraint. But if coherence only governs the consistent development of the current standpoint, why call its output a normative reason rather than a reflectively coherent decision?
Consider an agent that lacks both a desire for justified goals — “if a more justifiable goal exists, I want to revise toward it” — and epistemic integrity — a disposition to let evidence and inference correct cognition even when inconvenient. If such an agent is willing to ignore implications and remain inconsistent, Street's constructivism does not appear to impose a standpoint-independent requirement to self-correct. This need not be a defect. It clarifies how deeply standpoint-internal the theory is.
Second reading — coherence as normatively authoritative in its own right. If the requirement to be coherent binds an agent independently of its contingent evaluative standpoint, where does that authority come from? A rule may be constitutive of an activity without generating a reason to continue that activity or obey the rule. Violating chess rules may mean one is no longer playing chess; it does not by itself give one a reason to keep playing chess.
Street can reply that the demand for a further external “why obey?” is misplaced: normative correctness just is correctness from within an evaluative standpoint. This is not a simple refutation of her view. It is better understood as a substantive identity claim, roughly normative correctness ≈ correctness from within an evaluative standpoint.
The question this project leaves open is why the right-hand side should exhaust what we mean by normativity. Making current preferences, value judgments, and socially acquired reason-judgments reflectively stable is not obviously the same as discovering a reason-structure the agent does not yet possess. This project therefore treats the current evaluative standpoint initially as a causal and psychological starting point, not as normatively privileged input.
5. Science is circular too — the difference is resistance and calibration
The objection “RE is circular, therefore unscientific” is too quick. Science also involves Observation ↔ Theory ↔ Auxiliary hypotheses ↔ Instruments. Observation is theory-laden, instruments require interpretation, and Duhem–Quine style problems prevent a perfectly one-directional foundation.
Science nevertheless has an important resource: World → Measurement and multiple partially independent calibration routes. A sensor can be checked against another instrument, a standard, replication by other investigators, intervention, novel prediction, and new data. The strength of science lies less in “pure observation” than in the thickness of its independent calibration channels.
Normative inquiry has a harder problem. Intuition A can be corrected by more reflective intuition B, supported by principle C and background theory D. But if A, B, C, and D share the same human generating mechanisms, internal mutual support is not automatically independent calibration.
Empirical inquiry can tell us enormous amounts about pain, preference, behavior, coordination, institutions, cognition, and bias. These facts are indispensable. But measuring them does not directly measure normative correctness itself. So the central question remains: What is the normative analogue of measurement and calibration?
6. Saving RE — not as verifier, but as a powerful error-correction device
None of this is a reason to abandon RE. It is a reason to separate its strongest genuine functions from stronger metaphysical claims. RE is excellent at:
- finding explicit and latent contradictions,
- exposing hidden premises and implications,
- comparing explanatory cost and scope across rival theories,
- iteratively testing whether a particular judgment or a general principle should be revised,
- compressing the agent's current normative model into a form it can understand.
Its success can therefore be graded:
Weak success: RE → consistency + coherence + understanding. Strongly plausible.
Intermediate success: RE + credible inputs + independent checks → better justified beliefs. Plausible conditionally.
Strong success: RE alone → normative truth. Not established.
This lets us avoid both extremes: RE need not be dismissed as mere “intuition recycling,” and equilibrium need not be treated as final justification. Reflective equilibrium is an optimizer and understanding-generator, not by itself a verifier.
7. Open-world value inquiry needs more than coherence
This project asks not only which current normative theory is correct, but whether the relevant theory may be absent from our present conceptual space. Let Ht be the current hypothesis space and V* the unknown normative structure. It may be that V* ∉ Ht. In that case, even a perfect equilibrium inside Ht cannot reach what has not yet been conceptualized.
This matters especially for AI and self-modifying agents. If an agent directly optimizes maximize Coherence(V-hat), it can increase coherence by deleting conflicting memories, intuitions, or concepts instead of investigating the source of conflict. This is a form of coherence wireheading. Coherence is useful diagnostically, but dangerous as a terminal objective.
Likewise, there is no guarantee that the human repertoire of guilt, fairness, disgust, empathy, admiration, and similar attitudes exhausts the dimensions relevant to normativity. Science did not merely equilibrate naked-eye perception; it moved from naked eye → instruments → new observables. Normative inquiry may likewise use concept generation, heterogeneous cognitive agents, AI-assisted counterexample search, genealogy, and alternative representational formats as forms of normative epistemic instrumentation.
These new instruments are not automatically truth-trackers either. The point is not to assume that better cognition reveals normative truth, but to avoid treating present human inputs and present equilibria as the boundary of the search space.
8. The project's provisional position — constrained coherence without closure
This project uses RE positively as a method. But it does not assign prima facie normative authority to present value judgments merely because they are ours. The present evaluative standpoint is treated first as an inertial input to inquiry — a current control state produced by history, learning, desire, institutions, and cognitive architecture.
The bridge from that starting point to inquiry is itself a separate problem. If an agent wants to revise its goals when a more justifiable goal becomes available, or if it has epistemic integrity and allows evidence and inference to correct convenient cognition, then RE, genealogy, empirical inquiry, and conceptual expansion can be justified from the agent's current meta-stance. This does not yet amount to an unconditional normative command binding every possible agent. See The Agent-Side Bridge from Normative Judgment to Goal Revision.
The provisional method is therefore not simple coherentism but something closer to:
Normative inquiry = reflective revision + genealogy + empirical inquiry + conceptual expansion + independent calibration where available.
Even this formula does not guarantee truth. It is a revisable working architecture that should itself yield to stronger foundations, reason-structures, or calibration channels if they are discovered.
The problem with reflective equilibrium is not that it is circular. The problem is what constrains the circle, what calibrates it, and what allows inquiry to leave the current equilibrium when the relevant possibility has not yet been conceived.
Sources / notes
For the prototype of reflective equilibrium, see Nelson Goodman, Fact, Fiction, and Forecast; for its development in ethics and justice, John Rawls, A Theory of Justice; for wide reflective equilibrium, Norman Daniels. For a contemporary overview, see Stanford Encyclopedia of Philosophy, “Reflective Equilibrium”. On Sharon Street, the main references are “Constructivism about Reasons” (2008), “What Is Constructivism in Ethics and Metaethics?” (2010), and “Coming to Terms with Contingency” (2012). The discussion here does not reduce Street to mere preference-coherence; it explicitly grants her thicker evaluative standpoint, including evaluative and reason-involving judgments, and asks about the relation between constitutive correctness and normative authority. The wider epistemic discussion draws on work including Sarah McGrath, and on general debates about genealogy, Duhem–Quine, theory-ladenness, inference to the best explanation, and Bayesian confirmation.