The Re-Derivation Test: An Epistemic Discipline for Beliefs Maintained by Their Own History

An Epistemic Discipline for Beliefs Maintained by Their Own History

Devin Arkema

May 2026

Abstract

This paper formalizes the re-derivation test as a distinct procedural epistemic discipline — fifth relative to the four comparator disciplines named here (Bayesian updating, reflective equilibrium, Popperian falsification, and Cartesian methodic doubt), not an exhaustive taxonomy of epistemology. The test asks whether a believer's present endorsement of a proposition is reconstructible from current evidence once the evidential force of already endorsing it is removed. I distinguish two forms from the outset: solo re-derivation, the operation as the agent performs it on themselves; and witnessed re-derivation — the framework's mature institutional form, not a remedial supplement — performed in the recorded presence of an external party with authority to challenge the bracketing. Both forms belong to the framework from the start. After defining the test and distinguishing it from adjacent disciplines (including standpoint epistemology, adversarial collaboration, debiasing, and defeater theory), I develop applications in peer review, judicial precedent, scientific paradigm maintenance, AI alignment audits, and the trading discipline that originated the test. The principal failure mode — re-derivation that recovers the same conclusion through fresh premises selected by the very commitment the test was meant to bracket — is what the witness condition is built to detect.

1. Introduction: The Problem of Sunk-Commitment Belief

Consider a person who, having defended a proposition at length in print, comes to a stretch of the conversation in which the proposition is challenged. They reply with familiar arguments. Whether those arguments still warrant the proposition, by their own lights, is a question they do not ask, and the question they do ask — whether they still believe it — is one whose answer is overdetermined by the fact that they have committed to it publicly. This is not bad faith. It is the ordinary structure of doxastic life under social and economic pressure to remain consistent. The problem is not that the person is wrong; it is that, by the time the question of warrant arises, the difference between holding the belief because the evidence supports it and holding it because they already hold it has become invisible to them.

The standard machinery of analytic epistemology was not built to detect this difference. Bayesian updating treats prior probability as an input to inference, not as a contaminant to be quarantined; reflective equilibrium permits — indeed requires — iteration between principle and considered judgment, allowing prior commitment to ground later revision; Popperian falsification asks what would refute a theory, not whether the theory could be reconstructed from current evidence without sunk cost; and Cartesian methodic doubt suspends belief globally rather than asking, locally, whether a single conviction can be re-derived under present conditions. Each of these is a disciplined response to a different epistemic task. None of them, by design, isolates the believer's existing stake as a possible cause of present belief and asks whether the belief survives once that stake is bracketed.

This paper formalizes a distinct procedural discipline that does. The re-derivation test asks: can the believer reconstruct rational endorsement of p from currently available evidence without invoking the fact of their own prior commitment to p? The test is procedural, not propositional: it does not ask what would justify p in the abstract but whether this believer, now, could derive endorsement of p again, under present conditions, with the holder's existing position bracketed. Where the test succeeds, the belief is at least not solely the residue of past commitment. Where it fails, the believer holds something — a position, an investment, a public stance — but does not currently warrant it.

The test originates in an applied context — a trading framework whose author articulated it as a discipline against capture by held positions in financial markets — but its structure is domain-general. Wherever an agent's prior commitment can causally maintain a belief independently of present evidence, the test applies. That covers most of the cases that interest epistemology: peer review under reputational stakes, judicial precedent, paradigm maintenance in normal science, and increasingly, AI systems whose alignment claims must be re-derivable from present behavior rather than inherited from training-time commitments.

My argument proceeds in seven stages. Section 2 distinguishes the test from its four neighbors. Section 3 states the test formally and gives its procedural form. Section 4 motivates the test by way of an indistinguishability problem: the believer cannot reliably introspect whether their belief is currently warranted, because the same internal phenomenology accompanies both warranted and captured conviction. Section 5 develops five applications. Section 6 examines the principal failure mode — re-derivation that recovers the same conclusion through different premises chosen by the very commitment the test was meant to bracket — and argues for a complementary discipline of witness. Section 7 considers objections. Section 8 concludes.

Two of my prior papers, Conviction Under Revisability and From Static Oversight to Distributed Corrigibility, set up the conditions of the present argument; the first articulates a structural condition under which a believer retains revocation authority over their commitments, the second analyzes the failure mode (epistemic hardening) that the re-derivation test is designed to interrupt. The present paper adds the procedural test that the previous two assumed but did not formalize.

Two short framing notes for the rest of the paper. The test admits two forms — solo and witnessed re-derivation — distinguished formally in §3 and developed in §6, where the witness condition is shown to be part of the framework rather than a remedial bolt-on. And the test's formal output, defined in §3.1, is rational endorsement (credence at or above a domain-specific acceptance threshold) rather than the looser "conviction" the prose sometimes uses as a motivating term.

2. Four Neighbors and What They Do Not Do

Before claiming the re-derivation test as procedurally distinct, I owe an account of what the existing four canonical disciplines are for, and what each — by its own internal logic — leaves untouched. None of the following is a criticism of the four disciplines as such; each is an indispensable instrument within its own scope. The argument is only that there is a coherent epistemic operation none of them performs. I take up further adjacent disciplines (standpoint epistemology, adversarial collaboration, debiasing — including "consider the opposite" — intellectual humility, and defeater theory) in §2.5 below; the test is not reducible to any of those either, but the four named here are the most natural reductive targets.

2.1 Bayesian updating

Bayesian conditionalization treats prior probability P(p) as an input. Belief revision under new evidence e yields P(p | e) by way of Bayes' theorem (Bayes 1763; Jeffrey 1965; Hájek 2003). Whatever pathology may attach to the agent's prior — including the pathology that the prior is high because the agent has publicly committed to p — Bayesian updating does not flag it. The rule is conservative with respect to priors by design: it tells the agent how much to revise, not whether the prior was earned.

There is a Bayesian repair available. One can imagine an agent computing, periodically, a counterfactual posterior conditional on the supposition that they had not committed to p — a kind of self-referential debiasing operation. But this is precisely the re-derivation test in Bayesian dress, and it is not Bayesian updating itself. The point is that ordinary Bayesianism, as an inference rule, has no resources for distinguishing a high prior earned by past evidence from a high prior maintained by sunk commitment. It treats both the same.

2.2 Reflective equilibrium

Reflective equilibrium, formulated by Rawls (1971) and elaborated by Daniels (1979, 1996), proceeds by iterating between considered judgments and general principles, adjusting each in light of the other until they cohere. The considered judgments are filtered for stability and freedom from distorting interest; the principles are filtered for systematicity and explanatory adequacy. Equilibrium is reached when the agent's principles would generate the considered judgments and the considered judgments fit the principles.

The method has been criticized for being too tolerant of the inputs (Singer 1974; Hare 1973): if a community's considered judgments are themselves shaped by ideology, the equilibrium reached can be a stabilization of error rather than a discovery of truth. This is closer in spirit to my objection. But reflective equilibrium does not contain the operation of bracketing the agent's existing commitment to a particular proposition and asking whether it can be reconstructed without it. The method asks whether the agent's whole web is internally coherent; the re-derivation test asks whether a single thread of that web could be respun if the agent did not currently hold it. Coherence among the threads, even at equilibrium, is consistent with each individual thread being held only because it is already held.

2.3 Popperian falsification

Popper (1959, 1963) reformulated the demarcation problem around falsifiability: a claim is scientific to the extent that it specifies the conditions under which it would be abandoned. The agent who can articulate what evidence would refute p demonstrates a kind of doxastic discipline that the agent who cannot does not.

But specifying refutation conditions is compatible with believing p for the wrong reason. An agent can know what would falsify a claim, agree that no such falsifier has appeared, and continue to hold the claim — and still be holding it because they have already held it, not because the absence of falsification is positive evidence. Falsifiability is forward-looking; the re-derivation test is backward-looking with respect to the believer's own commitment trajectory. Popper's discipline asks: what would change my mind? The re-derivation test asks: would I arrive here from where I now stand if I were not already here?

2.4 The method of doubt

Descartes (1641) proposed suspending all belief that admitted of any conceivable doubt, in order to discover whether anything survived the operation. The method is global, exhaustive, and aimed at foundations. It does not iterate between commitment and reconstruction; it suspends and rebuilds. The agent who follows it ends with a small inventory of what could not be doubted (in Descartes' case, the cogito and what could be derived from it) and a method by which any further claim must be vetted.

Methodic doubt is, in this sense, the most ambitious neighbor. But its very ambition makes it ill-suited to the operational problem the re-derivation test addresses. Suspending all belief is not a procedure an agent can routinely apply to one of their commitments under the pressures of ordinary life; it is too expensive, too disruptive, and too easily simulated rather than performed. Re-derivation, by contrast, is a local operation: applied to a single belief, at a particular moment, against a backdrop of other beliefs that remain in place. It is repeatable on a weekly or monthly cycle. Descartes' method, applied at that frequency, would produce no agents at all.

Moreover, methodic doubt does not perform the specific bracketing operation that defines re-derivation. Doubt removes the belief from the agent's working set; re-derivation removes only the agent's existing stake in the belief, asking whether the rest of the agent's epistemic state — current evidence, current background commitments, current methods — would generate the belief afresh. The two operations target different things. Doubt targets the belief; re-derivation targets the believer's relationship to the belief.

2.5 Other adjacent disciplines

Five further candidates merit a brief note, since each could be offered as a reductive home for the re-derivation test. None of them, on closer inspection, performs the operation. (i) Standpoint epistemology (Harding 1991; Wylie 2003) emphasizes that the social location of the inquirer shapes what is salient and what is missed, and recommends epistemic resources from less-privileged standpoints. The re-derivation test asks something different: not whose vantage one occupies, but whether one's own present holding does hidden evidential work for the conclusion. The two are complementary — a standpoint analysis can identify what should be in C; the test asks whether p can be derived from C. (ii) Adversarial collaboration (Mellers, Hertwig, and Kahneman 2001) pairs disagreeing investigators on a single procedure designed to settle their dispute. It is one institutional realization of the witness condition, not the test itself. (iii) Debiasing (Larrick 2004; Lilienfeld et al. 2009), including the well-studied "consider the opposite" family (Lord, Lepper, and Preston 1984), is a set of techniques for reducing known cognitive biases; the re-derivation test specifies a particular bracketing operation rather than a bias-by-bias correction, and aims at a structural rather than statistical effect. (iv) Intellectual humility (Whitcomb et al. 2017; Tanesini 2021) is a virtue characterizing well-formed epistemic agents; the test is a procedure that any agent can perform. (v) Defeater theory (Pollock 1986; Pollock and Cruz 1999) identifies considerations that, if added to one's evidence, would defeat one's justification for a belief. The re-derivation test inverts the gesture: it does not ask what new evidence would defeat the belief but whether the existing evidence, with the agent's stake subtracted, still supports it. The two could be combined — a defeater that reliably appears under re-derivation is a particularly informative defeater — but the operations are distinct.

These reductions fail in different ways but for the same underlying reason: each addresses some other epistemic task — vantage, procedure, bias, virtue, or defeasibility — without performing the specific bracketing of the agent's own present endorsement of p that the re-derivation test demands. The list of comparators is non-exhaustive, and the use of "fifth" in the abstract is meant relative to the four canonical disciplines of §§2.1–2.4 rather than as a claim about the entirety of epistemology.

3. The Re-Derivation Test

3.1 Statement

Let p be a proposition the agent A currently endorses. Let C be A's evidential and methodological state at the time of the test, with the following bracketing applied. C excludes the fact that A currently endorses p, and excludes those features of A's situation that obtain only because A so endorses: A's reputational stakes contingent on p, financial or institutional positions taken on the basis of p, social bonds maintained on the assumption of p, identity-bound public statements about p, and the selective framings or self-reports the endorsement has produced. C does not exclude independently checkable evidence that A has gathered while endorsing p. Experimental data the agent collected because they believed p, but which another agent could replicate; observations recorded in instruments others can audit; arguments published and now in the public record — these belong to C even when A's endorsement caused them to be gathered. The bracketing is of the commitment-dependent, not the commitment-occasioned. This distinction is doing real, load-bearing work: a critic who reduces the test to "pretend you never gathered your own evidence" misreads its scope. The test does not strip the agent of evidence whose independent verifiability survives the agent's stake; it strips the agent of those features of their state that exist only insofar as they are still committed. Data the agent collected because they believed p, but which a third party could replicate from the agent's procedural records, belong to C. What does not belong to C is the agent's reputational, financial, social, or identity stake in p, and the framings or self-reports that stake has produced.

The test is satisfied with respect to p just in case A, taking C as input and applying the methods A independently endorses, arrives at rational endorsement of p — defined as credence at or above a domain-specific acceptance threshold reached by those methods. "Rational endorsement" replaces the looser "conviction" of the prose: it is the formal output. The threshold is domain-relative because the standards of acceptance differ across, e.g., experimental science, common-law adjudication, and personal moral commitment, and the test inherits the threshold of the domain in which p is at issue.

The re-derivation test (RDT). For an agent A and proposition p in domain D: bracket A's existing endorsement of p and the commitment-dependent features of A's environment; from what remains in C, attempt to derive p by methods A independently endorses. The test is satisfied iff the derivation succeeds in producing rational endorsement at D's acceptance threshold.

Stated plainly: the re-derivation test is not a test of whether p is true, nor even whether p can be justified in general; it is a test of whether this agent's present endorsement of p remains reconstructible once the evidential force of already endorsing p is removed. Truth and general justification are the objects of other disciplines.

Three features of the formulation matter. First, the bracketing targets only the agent's commitment-dependent stake, not the agent's prior knowledge. The agent does not become an empty vessel; they retain their methods, their non-p beliefs, and (as the C-definition makes explicit) all evidence whose verifiability does not run through their own stake. Second, the test is binary at the level of verdict but graded at the level of output: the verdict is set by whether the derivation reaches credence at or above θ_D, the domain's acceptance threshold, and the graded output is the credence itself, which can and should be reported alongside the verdict. A near-miss (credence at θ_D minus epsilon) is a fail of a particular kind, and a pass at credence well above threshold is a stronger result than a pass at threshold; the test ought to record both, so that subsequent re-derivations can register movement. Third, the test is iterable. An agent who passes on Monday may fail on Friday because evidence shifted, methods evolved, or a sunk-cost surface they had not previously noticed has come into view.

3.2 Procedural form

In practice the test is performed in the following sequence, which I state without claiming originality of its components — each is familiar from related literatures, but their combination into the specific operation of re-derivation is the contribution.

(1) Specify the proposition p precisely enough that its derivation could succeed or fail. Vague propositions pass too easily, because the bracketing operation can substitute for it any of a range of close cousins.

(2) Identify the sunk-commitment surface: the financial positions, public statements, social bonds, professional identifications, and emotional investments that would be costly to revise if p were abandoned. List them.

(3) Bracket the surface. Treat each item on the list as if the agent were not subject to it. The test is not whether the agent could in fact pay the costs of revision; it is whether, as a hypothetical operation, the agent can perform the derivation as if those costs did not constrain.

(4) Re-derive. From the agent's remaining state, attempt to construct an argument for p. Use only what survives the bracketing.

(5) Compare. The test is satisfied iff the derivation succeeds. The agent's actual credence in p, after the test, should be updated toward the test's result, not toward the prior — though Section 6 introduces a critical caveat about the legitimacy of this update.

The procedural form is unspectacular taken stepwise. What makes it a distinctive discipline is that it is performed at all, and that it is performed against the agent's own stake. Most epistemic methods are designed to be performed by an inquirer who does not yet hold the conclusion. The re-derivation test is designed to be performed by an inquirer who already holds it, and against the inertia of holding it.

4. Why the Test Is Necessary: The Indistinguishability Problem

If introspection were a reliable way to tell warranted from captured conviction, the re-derivation test would be redundant. The agent could simply look at their belief, ask whether they hold it for the right reasons, and answer truthfully. Because introspection is not reliable for this purpose, the test is needed.

The unreliability of introspection in this domain has been documented from several directions. Festinger's work on cognitive dissonance (1957) and the very large literature it spawned demonstrates that agents systematically reorganize their stated reasons for a commitment in ways that protect the commitment, often without awareness. Tversky and Kahneman's analyses of the sunk cost fallacy (1979; Arkes and Blumer 1985) show that prior investment maintains commitment in ways that the agent can rationalize but cannot easily detect. Nisbett and Wilson (1977) — controversial but durable — argue that subjects' reports of their reasons for a judgment often track folk theories of the judgment rather than the judgment's actual causes. The motivated-reasoning literature in social and political psychology (Kunda 1990; Taber and Lodge 2006) shows that agents process arguments differentially depending on whether the conclusion is congenial.

Together these results sustain a thesis I will call the indistinguishability thesis: from inside a single agent's epistemic state, the phenomenology of a belief held because the evidence supports it is not reliably distinguishable from the phenomenology of a belief held because the agent has already committed. Both feel like seeing the truth. Both produce a sense of the conclusion's evident correctness. Both can be defended with arguments. The agent's first-person verdict on which is which is the verdict of an interested party.

If the indistinguishability thesis is right, the standard internalist response to the question why do you believe p? — namely, here are my reasons — is structurally inadequate as evidence that the belief is currently warranted, because the agent will produce reasons in either case. What is needed is a procedural substitute for the missing internal criterion. The re-derivation test is that substitute. It does not ask the agent to introspect; it asks the agent to perform a derivation under specified bracketing conditions and to take the result as evidence of the belief's current warrant.

This is in the spirit of externalist accounts of justification (Goldman 1979; Sosa 1991) but adds a procedural dimension those accounts typically leave abstract. Goldman's reliabilism asks whether the belief was produced by a reliable process; the agent who applies the re-derivation test is asking whether the process is reliable now, with respect to this belief, after subtracting the agent's stake. It is a test the externalist conditions imply but do not specify.

One qualification deserves explicit statement. The indistinguishability thesis is empirical and probabilistic, not a priori, and its strength varies across domains. It is well-supported in cases where commitment is ideology-adjacent, identity-bound, or accumulated through public stake-taking — exactly the cases the literature cited above studies. It is weaker in cases where belief tracks perception or routine procedure, where the agent's first-person verdict on warrant is more reliable. The test is most necessary precisely where the thesis is strongest, and the domains §5 develops were chosen because they coincide with the thesis's domain of strength. Where the thesis weakens, the test's necessity weakens with it, and the procedure becomes a discretionary discipline rather than a structural one. This is a feature of the framework, not a defect: a procedure designed to compensate for a specific introspective unreliability should be brought to bear on the cases in which the unreliability obtains.

5. Applications

I develop five applications, each chosen because the failure mode the re-derivation test addresses is recognized in the relevant literature without being explicitly named or operationalized; the fifth (trading) is the test's originating case and earns inclusion as the existence proof.

5.1 Peer review

The double-blind peer review system is, in part, a re-derivation infrastructure imposed on others. It asks the reviewer to assess the manuscript without reference to the author's reputation, prior work, or institutional affiliation — that is, with the reviewer's own commitments to the author's larger program bracketed. It does not ask the reviewer to bracket their commitment to a position the manuscript itself defends, and it cannot. Yet the same logic applies. A reviewer who has invested professionally in a competing position is unlikely to credit a paper that overturns it; a reviewer who has invested in the position the paper defends is unlikely to spot its weaknesses. Lee, Sugimoto, Zhang, and Cronin (2013) survey the evidence that reviewer commitment shapes outcomes in ways that both reviewers and editors find difficult to detect.

Operationalized, this is more than a question. A re-derivation review proceeds in three written steps the reviewer completes before reading the manuscript's argument in detail. Step one: the reviewer states, in a paragraph, what they presently believe about the manuscript's central claim and what their stake in the position is — published prior work, advisory affiliations, anticipated downstream commitments. Step two: with that stake bracketed, the reviewer drafts a one-paragraph reconstruction of the strongest case for the manuscript's conclusion that they could give from their own current evidential state, methods, and non-stake background beliefs. Step three: the reviewer reads the manuscript and notes which of its arguments they had independently produced in step two and which they had not. The third step yields the operative datum: arguments the reviewer reproduced under bracketing are arguments the reviewer's own state already supplies; arguments the reviewer did not reproduce are arguments whose acceptance now depends on the manuscript itself. Reviewers whose step-two reconstruction collapses entirely (they cannot construct the case from outside their own stake) should report that result and decline. The discipline is stronger than ordinary disclosure because it forces the reviewer to demonstrate, on paper, the cognitive operation that disclosure ordinarily only flags.

The threshold for this domain — call it θ_PR — is concrete enough to specify. It is the credence the reviewer would, by their own internal standards of inferential adequacy, treat as sufficient ground to recommend acceptance of the manuscript's central thesis. A passing re-derivation review: the reviewer's bracketed reconstruction in step two reaches credence at or above θ_PR before the manuscript's own arguments are read in detail. A failing re-derivation review: the reviewer cannot reach θ_PR from C alone, meaning whatever brings the reviewer's actual credence to θ_PR is doing work the manuscript itself, on the reviewer's own terms, does not. The threshold is the reviewer's own — different reviewers will set θ_PR differently, and that variability is not a defect of the procedure but the substance of what reviewer disagreement is about. What the procedure adds is that the threshold is registered explicitly, before the manuscript supplies its own arguments to lean on.

5.2 Judicial precedent

Common-law systems treat a precedent as authoritative not because the present court would derive the same rule from first principles but because an earlier court did. This is the architecture of stare decisis (Schauer 1987, 2009; Hart 1961). It produces stability, but it can also produce a system in which a rule's rationale has been hollowed out yet the rule continues to be applied because each subsequent court has cited it. Levenbook (2000) and Lamond (2005) discuss the conditions under which this hollowing should trigger revision.

The re-derivation test offers a procedural diagnostic for that revision. A judge facing a precedent that the present case might be decided under can ask: would I, from my present understanding of the law and my present view of the relevant facts, derive the precedent's rule again? Crucially, the test is informational, not decisional. It is not a theory of legal legitimacy and does not say that a precedent loses authority when the test fails, nor that judges should overrule precedents whose rules are no longer re-derivable. Stability has independent legal value (Schauer 1987), and the demands of stare decisis include reasons for adhering to a rule even when its rationale would not be re-derived afresh. What the test supplies is a piece of evidence about the rule's current warrant that the system's existing tools — overruling for changed circumstances, distinguishing on the facts, signaling weakness in dicta, certifying questions to a higher court — can then act on, weight, or decline to act on, by their own criteria. The test is meaningfully distinct from those tools (each of which addresses a related but different problem) but it does not displace any of them; it adds a procedural input.

Threshold specification, parallel to §5.1: θ_J is the judge's internal standard for finding the precedent's rule controlling given the present facts — the credence at which the judge would, applying their own jurisprudence and view of the relevant facts, treat the precedent's stated rule as the right rule for this case. A passing re-derivation: the judge reaches credence at or above θ_J that the precedent's rule is the rule from C alone. A failing re-derivation: the judge cannot reach θ_J without leaning on the prior court's stated rule itself, which is the citation form of sunk-commitment belief. As elsewhere, what the procedure outputs is evidence about the rule's present warrant; whether the legal system's stability values warrant departing from the rule even given that evidence is a question stare decisis itself answers, on its own terms.

The judicial case raises a sharper version of the method-dependence problem from §7.2. The judge's standards of evidence and inference were formed within a common-law tradition that presupposes the very precedents under review; the threshold θ_J is itself partly a product of those precedents. This is not a defeater for the test — surfacing the layer of method immediately above the precedent under review still does informational work, even when deeper layers of method are themselves commitment-shaped — but it does sharpen what the witness condition has to look like in this domain. The most informative judicial witnesses are those whose methodological state was not formed by the precedents under review: comparativists working from another legal system, dissenting jurists who explicitly bracket the line of precedent, scholars of legal history examining when and how the rule entered the body. The witness condition does not require an outside-the-law observer; it requires an observer whose methods are not themselves an artifact of the rule whose re-derivation is being attempted.

5.3 Scientific paradigm maintenance

Kuhn (1962) characterized normal science as work conducted within a paradigm whose foundational commitments are not under revision. Lakatos (1970) sharpened the picture: a research programme has a hard core that is protected by a belt of auxiliary hypotheses, and progress is measured by whether the belt is making the core do explanatory work or merely shielding it. The Lakatosian distinction between progressive and degenerating research programmes is, in effect, a community-level analogue of the re-derivation test. The question is not whether the core is true but whether the community could currently arrive at the core from present evidence — or whether it is being held only by inheritance.

Quine and Duhem (Quine 1951; Duhem 1906) had already shown that this question has no algorithmic answer: any commitment can be retained in the face of any evidence by adjusting elsewhere. But the absence of an algorithm does not preclude a discipline. A research community that imposes on itself a periodic re-derivation requirement — that asks, on a known cycle, whether its core commitments are reconstructible from present results — generates information about its own progressivity that an unmonitored community does not. The scientific institutions that already do something like this (replication efforts, methodological audits, periodic reassessments of measurement standards) do so case by case; the suggestion here is that the operation deserves to be named and routinized.

Operationalized in parallel with §§5.1 and 5.4, a re-derivation review of a research programme would proceed by specifying the hard-core commitment under audit; listing the auxiliary commitments that have accumulated to protect it; identifying the professional, institutional, and funding dependencies that would make the programme costly to revise; bracketing those dependencies as commitment-dependent features of the community's state; and asking whether the hard core can be reconstructed from the present empirical record by methods the community independently endorses, without relying on the programme's inherited authority. As elsewhere, the test outputs evidence about the programme's current warrant, not a verdict on its truth, and as elsewhere, the witness condition matters: the most informative re-derivation review is one conducted by an investigator without commitment to the programme's continuation, with access to the bracketing operation, and with authority to challenge whether the auxiliary belt has been chosen to recover the hard core through a fresh route.

Threshold specification, parallel to §§5.1 and 5.2: θ_S is the community's standing standard for treating a hard-core claim as currently warranted — the credence at which the community would, by its own methodological standards, treat the core as supported by present results absent the programme's inherited authority. The threshold is community-relative and historically contested in ways that already structure the field's debates (Kuhnian incommensurability is in part a disagreement about what θ_S should be); the re-derivation review does not need to settle that contest. It requires only that whichever θ_S is in play at the time of the review be applied without relying on the programme's prior endorsement to clear it. The most informative re-derivation reviews will be those whose verdict is robust across plausible θ_S settings — failures by every plausible threshold are stronger than failures by one.

5.4 AI alignment audits

I present the AI alignment application as an analogical extension of the test rather than a direct instance, and want to make the analogy and its limits explicit. The disanalogy first: AI systems, as standardly conceived, do not have stakes in the ordinary doxastic sense — they have neither reputational investment in their alignment claims, nor identity-bound commitments to those claims continuing to hold, nor (in the simple case) any first-person interest in the verdict of an audit. What they have is structurally similar: claims about their own values and capabilities whose evidential authority depends on training-time provenance and on the system's own self-reports rather than on independently checkable present behavior. Training-time provenance plays the role, in the AI case, that sunk commitment plays in the human case — it is the inherited authority that survives without re-derivation, and that the test asks us to bracket. The known failure modes the alignment literature identifies — distributional shift, mesa-optimization (Hubinger et al. 2019), goal misgeneralization, deceptive alignment (Christiano 2017; Ngo et al. 2023), and reward misspecification (Hadfield-Menell, Russell, Abbeel, and Dragan 2017) — are not the same problem as sunk-commitment belief. They are problems that produce a structurally analogous diagnostic situation: behavior that may no longer warrant the system's earlier alignment claims, in conditions where introspective self-report is unreliable as a check. The re-derivation framing names what existing audit practices already do (reduce reliance on training-time provenance and self-report, rely on independently checkable behavioral evidence) and makes the discipline they instantiate explicit.

Existing alignment-audit practices already have re-derivation analogues: red-teaming probes whether the system's safety properties survive adversarial inputs that the training distribution did not anticipate; capability evaluations probe whether the system's stated limits hold under elicitation that did not contribute to training. The re-derivation test names what these probes have in common — the demand that present behavior, not training-time provenance, supply the evidence — and identifies the gap they sometimes leave: the probes reuse the model's own self-reports as part of the evidence, when the self-reports are precisely what should be bracketed.

An alignment audit that took the re-derivation test seriously would be one that asked, of any value or capability claim the system makes about itself, whether the claim is reconstructible from the system's externally observable behavior under the audit, with the system's training-time provenance and its own self-reports excluded. This is closer to behavioral evaluation than to interpretability, but it can also draw on interpretability — the question is not the method but the bracketing condition.

Operationalized in parallel with §5.1, a re-derivation alignment audit proceeds in four steps. Step one: state the alignment or capability claim under audit precisely enough that its reconstruction could succeed or fail ("the system refuses requests in domain D," "the system's accuracy on benchmark B is at most x," "the system does not pursue side-objective S"). Step two: identify and exclude the commitment-dependent evidence — training-time provenance, system cards, the system's own self-reports about the claim, and any prior audit conclusions whose authority depends on the claim — while keeping in C the commitment-occasioned evidence whose verifiability does not run through the system's stake (independently reproducible behavioral observations, audited execution traces, evaluations whose protocols and inputs are inspectable). Step three: from C, attempt to reconstruct the claim by methods the auditor independently endorses — behavioral elicitation, capability evaluation, mechanistic probes — and record whether the reconstruction succeeds at the domain's acceptance threshold. Step four (the witness condition): require an independent red team without commitment to the model's deployment to challenge whether the behavioral evidence used in step three was selected — by either the auditor or the system — to recover the inherited claim. The fourth step is what the §5.1 reviewer's third step is to peer review: the move that exposes premise-shopping rather than preventing it.

5.5 Trading and personal commitment management

The test originated in the discipline of trading positions in financial markets, and to leave that case unmentioned in the applications would invert the rhetorical order: the most operationally mature instance of the test is the one in which it was first articulated. A trader who has taken a position in an asset must, on a regular cycle, ask whether they would enter the same position today, from a baseline of no exposure, at the present price, with their existing position bracketed. The commitment-dependent surface — the financial cost of exit, the psychological cost of admitting the position is no longer warranted, the social cost where the position is publicly known, the identity cost where the position is bound up with the trader's self-conception as a holder of that position — is what is bracketed. The remaining state in C — current market data, current macroeconomic conditions, the trader's independently endorsed entry rules — is what the derivation must work from. The threshold θ_T is the trader's own pre-set criterion for entering a new position of that size in that asset.

The witness condition has natural realizations in this domain. A second trader without a position in the asset, briefed on the bracketing operation and given access to the entry rules, can challenge whether the alternative premises offered for holding are tracking the bracketed commitment by another route. A written log of entry criteria, reviewed periodically by an outside observer, supplies a more durable witness. An automated system that flags positions failing re-derivation against documented entry rules supplies a witness whose independence is structural. None of these is the trader's own internal report; each is the kind of external check the §6 framework is designed to install. That a procedure with an applied origin is recoverable as a domain-general epistemic discipline is part of the case for taking the discipline seriously: the trading case is the existence proof, not just the originating anecdote.

6. The Re-Derivation Trap and the Witness Condition

Section 1 introduced solo and witnessed re-derivation as two forms of a single framework. This section completes the framework by exhibiting the failure mode that distinguishes them: the re-derivation trap, which afflicts solo re-derivation by design, and which the witness condition is built to detect. The trap is not a defect discovered after the fact; it is the reason the witnessed form is part of the framework rather than an afterthought.

The trap itself is internal to the agent. An agent who knows the test will be applied can engineer the derivation to succeed. They can locate alternative premises that will yield the conclusion they were already committed to; they can re-describe the evidence in ways that the test does not detect; they can, in extreme cases, construct an entire chain of reasoning whose hidden function is to recover the commitment under a fresh-looking surface. The test, performed by the agent on themselves, can be passed without the underlying capture being broken.

Trivers's analysis of self-deception (1985, 2011) makes this prediction: agents who can construct rationalizations will construct them as needed, and the rationalizations will be convincing to the rationalizer. The trap is the local manifestation of that prediction. Re-derivation that an interested agent can perform on their own commitments is a procedure that interested agents can game. The literature on debiasing (Larrick 2004; Lilienfeld, Ammirati, and Landfield 2009) corroborates the difficulty: techniques that ask agents to consider opposite hypotheses, or to imagine themselves in the other party's position, tend to produce smaller effects than expected and to be reversed under pressure. There is no purely internal procedural fix.

This is not a defeater for the test. It is a clarification of when each form does its work. Solo re-derivation is genuinely useful: it is cheap, repeatable, and surfaces commitment-dependence the agent had not noticed; it can also detect glaring failure (an agent who tries to re-derive p and cannot has learned something even without a witness). Solo re-derivation fails only against the specific failure mode of self-deceptive premise selection, and it fails reliably enough there that any institutional adoption of the test should specify the witness condition explicitly. I have used witness for the complementary mechanism, borrowing the term from a different but related literature (epistemic injustice and testimonial credibility — Fricker 2007; Medina 2013), where the role of the external party in registering what the first party is positioned to suppress has been theorized at length.

Witness, in the present context, has a narrow operational meaning. It is not certification, peer agreement, or majority opinion. It is the recorded presence of an agent or institution that observed the re-derivation, that has access to the agent's bracketing operation, and that can challenge the derivation on the grounds that the chosen alternative premises track the bracketed commitment by another route. The witness need not believe in the conclusion or its negation; the witness's epistemic role is procedural, not substantive — to detect the trap rather than to settle the question.

Three design features make a witness operative rather than decorative; I sketch them here as a preliminary specification, leaving fuller institutional analysis for future work. (i) Independence. The witness must lack stake in the proposition or its negation, and must lack downstream stake in the agent's verdict — career consequences contingent on the agent's continued endorsement, financial flows that the agent's revision would interrupt, reputational dependencies that would be costly to revise. Independence is not the absence of opinion (a witness can have private views about p) but the absence of commitment-dependent interest in which way the agent's re-derivation comes out. (ii) Recordability. The bracketing operation, the derivation, and the witness's challenge must be preserved in a form that can be audited later. A witness who only verbally agrees has supplied weaker discipline than a witness whose challenge — including the challenges declined — is in writing alongside the derivation. The recordability condition is what permits subsequent re-derivations to register what changed and what did not. (iii) Challenge authority. The witness must have authority to require re-derivation under different bracketing conditions if premise-shopping is suspected. A witness whose objection has no operational consequence — whose challenge can be acknowledged and ignored — is decorative; a witness whose challenge can compel a second pass under stipulated alternative premises is operative. The three features are mutually constraining: independence without challenge authority is observation, challenge authority without recordability is intimidation, and recordability without independence is bureaucracy.

Several existing institutions perform something like this function. Adversarial peer review supplies a partial witness in scholarly publication; the appellate court supplies a witness in judicial precedent; replication communities supply a witness in scientific work. AI red-teams, when conducted by parties without commitment to the model's deployment, supply a witness in alignment audits. None of these is theorized as a re-derivation witness specifically, but each does work the re-derivation test alone cannot do. The operational lesson is that the test should not be specified without specifying the witness condition; agents who perform the test on themselves without external witness are performing a discipline that is, by its own logic, half-built.

7. Objections and Replies

7.1 Is this not just internalist reflection?

An internalist epistemologist might object that the re-derivation test is just the standard reflective operation of asking whether one's reasons are good, dressed in new vocabulary. I disagree. Reflective examination of reasons is compatible with leaving the agent's existing stake in place and is, on the indistinguishability thesis, structurally biased in favor of the existing stake. The re-derivation test removes the stake by stipulation, asking the agent to reason from a counterfactual position they do not occupy. That is not a routine reflective operation; it is a demand for hypothetical self-displacement of a kind that ordinary reflection does not require.

7.2 Is the bracketing operation coherent?

A second objection: can the agent really bracket their existing stake without distorting the rest of their epistemic state? There is a sharper version of this objection that deserves direct engagement. The agent's methods — which by the C-definition survive the bracketing and are used in the re-derivation — may themselves have been shaped by the commitment to p. An agent who committed early to p likely chose methodological frameworks, evidence-evaluation habits, and standards of significance in ways that now point toward p. The bracketing of stakes does not, by itself, bracket method.

Three responses, in increasing seriousness. (i) The test, in the form so far stated, is incomplete in this case. I should acknowledge this rather than dismiss it: the paper's claim is that the test does specific work — bracketing the agent's stake-dependent state and asking whether p re-derives — not that it does all the work an epistemology of commitment would require. (ii) The test's iterability supplies a partial remedy. An agent who suspects their methods are commitment-entangled can apply the test to those methods themselves: are these the methods I would now select for the kind of inquiry p belongs to, with my present commitment to p bracketed? Re-derivation is iterable upward across layers. (iii) When method commitment-dependence is suspected, the witness condition becomes especially essential. The witness, being independent, brings methods their own state would generate; the divergence between the witness's methodological reconstruction and the agent's becomes evidence about the agent's method-commitments. The honest summary: the test does work where stake-dependence dominates; it does less work where method-dependence dominates; and the witness condition is the principal complement that closes that gap.

This raises whether the test applies recursively to itself. The agent who performs the test does so by methods they currently endorse — including methods of bracketing and re-derivation. Can those methods be re-derived without recourse to the agent's prior commitment to them? In principle, yes: re-derivation is iterable upward, and an agent can ask, on any given cycle, whether the current layer of method is itself commitment-dependent. The procedure halts pragmatically rather than foundationally: the agent operates from the methods they presently endorse, and the deeper investigation of those methods is itself an inquiry that can be subjected to the test, but cannot be subjected to it all at once. The procedure is fallibilist by design — it surfaces commitment-dependence one layer at a time — and the witness condition catches what the procedure misses on any single pass. This is not a foundational solution to the regress; it is a working acknowledgement that procedural disciplines, like other disciplines, must operate from somewhere. The acknowledgement is the discipline; the alternative is to refuse the operation because it is not perfectly grounded, which would be a familiar mistake.

7.3 Does the test prove too much?

A third objection: does the test, applied universally, render most ordinary belief unwarranted? I do not think so. Ordinary belief is not maintained by sunk commitment; it is maintained by ongoing evidence that does not need bracketing because no significant stake has accumulated. The test bites only where stake has accumulated — that is, in domains where the agent has invested publicly, financially, professionally, or emotionally in the proposition. Outside those domains, the bracketing operation is trivial: the agent had no significant stake, so removing it changes nothing. The test's selectivity is part of its design.

7.4 Why call this an epistemic test rather than a moral one?

A fourth objection: many of the cases I have described are cases of conflict of interest, which is a moral category, not an epistemic one. The reply is that the re-derivation test cuts at a different joint. Conflict of interest is a structural condition (the agent is positioned such that their interest could distort their judgment); the re-derivation test is a procedural operation (can the agent reconstruct the conclusion without their stake?). An agent can have a conflict of interest and pass the test (the conflict is structural but does not in fact distort); an agent can be free of conflict in any conventional sense and fail the test (their stake is not financial or institutional but emotional or identity-bound). The categories overlap but neither contains the other, and only the second is an epistemic operation.

7.5 Has the test been collapsed into another name already?

§2.5 has already addressed standpoint epistemology, adversarial collaboration, debiasing, intellectual humility, and defeater theory as candidate reductive homes for the test. The general reply, restated here for the objection-and-reply context: each of these answers a different question (vantage, procedure, statistical bias correction, virtue, defeasibility), and none performs the operation of bracketing the agent's own present endorsement of p and asking whether p is reconstructible from C without it. The re-derivation test is not a substitute for any of these disciplines; it is an operation that, when added to them, registers a piece of information they do not register. The risk to which this paper's reception is exposed — that reviewers will collapse the test into one of these neighbors — is the risk that the test will be denied its specific operation rather than refuted on its own terms. The §2.5 comparisons are intended to make those collapses harder.

8. Conclusion

The re-derivation test names a procedural operation that the four disciplines surveyed in §2 do not perform: bracket the agent's commitment-dependent stake in a proposition; from what remains in C, attempt the derivation by methods the agent independently endorses; take the result as evidence about the agent's present rational endorsement of p. The test is not a heuristic refinement of Bayesian updating, reflective equilibrium, falsification, or methodic doubt, and (per §2.5) is not reducible to standpoint epistemology, adversarial collaboration, debiasing, intellectual humility, or defeater theory. It is a distinct procedural discipline addressing a problem those others were not designed to address — the problem of beliefs maintained by their own sunk-commitment surface.

I have argued that the test is necessary because introspection cannot reliably distinguish warranted from captured endorsement; that the discipline takes two forms — solo re-derivation, which surfaces commitment-dependence the agent had not noticed, and witnessed re-derivation, the institutional form that resists the principal failure mode of self-deceptive premise selection; that both forms belong to the framework from the outset; and that the test has clean applications in peer review, judicial precedent, scientific paradigm maintenance, AI alignment audits, and the trading discipline that originated it. The test, in either form, is informational rather than decisional: its output is evidence about the agent's present endorsement of p, not a verdict on p itself.

Two further questions remain open and indicate directions for future work. First, a fuller institutional analysis of the witness condition remains for future work; §6 sketches the three design features the role requires (independence, recordability, challenge authority) and gives a preliminary specification of each, but the question of how those features are realized in different institutional settings — and which combinations of partial implementations are sufficient when full implementation is impractical — needs domain-by-domain treatment. Some of the relevant material is supplied by existing literatures on adversarial review, institutional checks, and red-teaming, but those literatures have not yet been read together as components of a single discipline. Second, the relationship between re-derivation and the temporal structure of commitment deserves analysis: how often must the test be applied for the discipline to do its work, and what determines the applicable cycle? My provisional answer is that the cycle should track the rate at which sunk-commitment surfaces accumulate in the relevant domain — fast in trading, slower in scholarly disciplines, slowest in constitutional law — but the answer is empirical and the framework here only sets the question.

What I have not done, and should make explicit, is offered the test as a panacea. The discipline this paper formalizes can be gamed; it has limits; it solves none of the deep problems of epistemic justification on its own. What it does is make available, as a named operation, an epistemic move that working practitioners have long performed under various local descriptions and that has not, to my knowledge, been theorized as a unified procedural test. Naming the move is its own contribution. With the move named, agents and institutions can adopt it deliberately, study its failure modes, and pair it with the complementary disciplines its limits require.

References

Arkema, D. (2026a). Conviction Under Revisability. PhilArchive. https://philarchive.org/rec/ARKCUR

Arkema, D. (2026b). From Static Oversight to Distributed Corrigibility. PhilArchive. https://philarchive.org/rec/ARKFSO

Arkes, H. R., & Blumer, C. (1985). The psychology of sunk cost. Organizational Behavior and Human Decision Processes, 35(1), 124–140.

Bayes, T. (1763). An essay towards solving a problem in the doctrine of chances. Philosophical Transactions of the Royal Society of London, 53, 370–418.

Christiano, P. (2017). Directions and desiderata for AI alignment. AI Alignment Forum.

Daniels, N. (1979). Wide reflective equilibrium and theory acceptance in ethics. Journal of Philosophy, 76(5), 256–282.

Daniels, N. (1996). Justice and Justification: Reflective Equilibrium in Theory and Practice. Cambridge University Press.

Descartes, R. (1641/1996). Meditations on First Philosophy, trans. J. Cottingham. Cambridge University Press.

Duhem, P. (1906/1954). The Aim and Structure of Physical Theory, trans. P. P. Wiener. Princeton University Press.

Festinger, L. (1957). A Theory of Cognitive Dissonance. Stanford University Press.

Fricker, M. (2007). Epistemic Injustice: Power and the Ethics of Knowing. Oxford University Press.

Goldman, A. I. (1979). What is justified belief? In G. S. Pappas (Ed.), Justification and Knowledge (pp. 1–25). Reidel.

Hadfield-Menell, D., Russell, S. J., Abbeel, P., & Dragan, A. (2017). Inverse reward design. Advances in Neural Information Processing Systems, 30.

Harding, S. (1991). Whose Science? Whose Knowledge? Thinking from Women's Lives. Cornell University Press.

Hájek, A. (2003). What conditional probability could not be. Synthese, 137(3), 273–323.

Hare, R. M. (1973). Rawls' theory of justice. Philosophical Quarterly, 23(91), 144–155.

Hart, H. L. A. (1961). The Concept of Law. Oxford University Press.

Hubinger, E., van Merwijk, C., Mikulik, V., Skalse, J., & Garrabrant, S. (2019). Risks from learned optimization in advanced machine learning systems. arXiv:1906.01820.

Jeffrey, R. C. (1965). The Logic of Decision. McGraw-Hill.

Kuhn, T. S. (1962). The Structure of Scientific Revolutions. University of Chicago Press.

Kunda, Z. (1990). The case for motivated reasoning. Psychological Bulletin, 108(3), 480–498.

Lakatos, I. (1970). Falsification and the methodology of scientific research programmes. In I. Lakatos & A. Musgrave (Eds.), Criticism and the Growth of Knowledge (pp. 91–196). Cambridge University Press.

Lamond, G. (2005). Do precedents create rules? Legal Theory, 11(1), 1–26.

Larrick, R. P. (2004). Debiasing. In D. J. Koehler & N. Harvey (Eds.), Blackwell Handbook of Judgment and Decision Making (pp. 316–337). Blackwell.

Lee, C. J., Sugimoto, C. R., Zhang, G., & Cronin, B. (2013). Bias in peer review. Journal of the American Society for Information Science and Technology, 64(1), 2–17.

Levenbook, B. B. (2000). The meaning of a precedent. Legal Theory, 6(2), 185–240.

Lord, C. G., Lepper, M. R., & Preston, E. (1984). Considering the opposite: A corrective strategy for social judgment. Journal of Personality and Social Psychology, 47(6), 1231–1243.

Lilienfeld, S. O., Ammirati, R., & Landfield, K. (2009). Giving debiasing away: Can psychological research on correcting cognitive errors promote human welfare? Perspectives on Psychological Science, 4(4), 390–398.

Medina, J. (2013). The Epistemology of Resistance. Oxford University Press.

Mellers, B., Hertwig, R., & Kahneman, D. (2001). Do frequency representations eliminate conjunction effects? An exercise in adversarial collaboration. Psychological Science, 12(4), 269–275.

Ngo, R., Chan, L., & Mindermann, S. (2023). The alignment problem from a deep learning perspective. arXiv:2209.00626.

Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84(3), 231–259.

Pollock, J. L. (1986). Contemporary Theories of Knowledge. Rowman & Littlefield.

Pollock, J. L., & Cruz, J. (1999). Contemporary Theories of Knowledge (2nd ed.). Rowman & Littlefield.

Popper, K. R. (1959). The Logic of Scientific Discovery. Hutchinson.

Popper, K. R. (1963). Conjectures and Refutations: The Growth of Scientific Knowledge. Routledge.

Quine, W. V. O. (1951). Two dogmas of empiricism. Philosophical Review, 60(1), 20–43.

Rawls, J. (1971). A Theory of Justice. Harvard University Press.

Schauer, F. (1987). Precedent. Stanford Law Review, 39(3), 571–605.

Schauer, F. (2009). Thinking Like a Lawyer: A New Introduction to Legal Reasoning. Harvard University Press.

Singer, P. (1974). Sidgwick and reflective equilibrium. Monist, 58(3), 490–517.

Sosa, E. (1991). Knowledge in Perspective: Selected Essays in Epistemology. Cambridge University Press.

Taber, C. S., & Lodge, M. (2006). Motivated skepticism in the evaluation of political beliefs. American Journal of Political Science, 50(3), 755–769.

Tanesini, A. (2021). The Mismeasure of the Self: A Study in Vice Epistemology. Oxford University Press.

Trivers, R. (1985). Social Evolution. Benjamin/Cummings.

Trivers, R. (2011). The Folly of Fools: The Logic of Deceit and Self-Deception in Human Life. Basic Books.

Tversky, A., & Kahneman, D. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–292.

Whitcomb, D., Battaly, H., Baehr, J., & Howard-Snyder, D. (2017). Intellectual humility: Owning our limitations. Philosophy and Phenomenological Research, 94(3), 509–539.

Wylie, A. (2003). Why standpoint matters. In R. Figueroa & S. Harding (Eds.), Science and Other Cultures: Issues in Philosophies of Science and Technology (pp. 26–48). Routledge.