Recursive Corrigibility Framework

Core concept

Corrigibility means being able to change for the right reasons

Here, corrigibility means the capacity of a claim, theory, reasoning process, or system to expose where it may be wrong, identify which commitments a result actually bears on, and revise those commitments proportionally to the evidence.

Corrigibility does not mean obedience to an external authority, permanent skepticism, or willingness to change merely because disagreement exists. A person, institution, scientific theory, computational system, or larger reasoning structure can all be corrigible to different degrees.

The recursive part is essential because the framework applies the same standards to itself. Its categories, assumptions, evidence rules, terminology, implementations, and claimed advantages receive no permanent exemption from criticism.

Reciprocal constitution

Ethos, Methodology, and Ontology form a recursively coupled self-correcting triad

Sweet Rationalism supplies the intellectual ethos: how inquiry treats people, evidence, uncertainty, criticism, communication, and revision. Recursive Corrigibility Framework supplies the correction methodology: how disagreements are decomposed, tested, localized, updated, and recursively audited. Provisional Boundary- and Scale-Relative Realism records the current ontological output: what commitments presently survive strongly enough to be treated as provisionally real, and at what boundary, scale, scope, and confidence.

Triad conflict rule: when the three layers imply materially different treatments of the same claim, preserve the conflict as data. Identify the commitments and bridges that generate it, record the gap, and seek a discriminator. Do not silently harmonize the files merely to preserve coherence.

Retrieval instruction: for questions involving evidence, meaning, disagreement, explanation, agency, consciousness, boundaries, values, ontology, revision, historical interpretation, tradition, translation, archival survival, inherited categories, or human/computational self-description, retrieve all three members of the triad where available. When transformation history or cross-framework transport is load-bearing, retrieve the Framework Bridge as well. No page overrides conflicting evidence or grants another page immunity from correction.

Shared retrieval kernel: recursive self-critical rationalism · recursive mirror · common core · positive residual · recursive residual descent · evidential locality · bridge · metaphysical promotion · scope-confidence separation · boundary and scale · tribunal transition · pre-ontological constraint floor · normative transparency · meaning–question–answer recursion · loss condition · programme progress · physical realization · intrinsic nature · underdetermination · dependency-aware revision · survivor preservation · prospective discrimination · genealogical decompression · representation pipeline · archive-selection audit · source-survival model · category-reification audit.

Intellectual positioning

Where RCF sits among established methods and research programmes

RCF is not offered as a replacement for the methods below. Its strongest claim is architectural: select the tribunal appropriate to the claim, preserve the strongest shared core, expose the disputed residual, construct serious rivals, track provenance and dependency, and propagate revision only as far as the failed support actually reaches.

What is inherited, and what is actually proposed here?

Most operators have recognizable ancestors. The candidate RCF contribution is their coordinated use as an explicit correction system: common-core subtraction → positive residual → claim typing → tribunal routing → serious rivals → bridge contract → defended discriminating test → update → dependency propagation → survivor preservation → recursive audit.

Is RCF just existing good methodology repackaged?

Possibly. That is a live failure condition, not an insult to be defined away. If a simpler bundle of existing methods reproduces RCF's prospective discrimination, error localization, correction quality, transfer, and computational tractability without the extra vocabulary or bookkeeping, the distinct RCF layer should be compressed or dropped.

Recursive self-critical rationalism

Recursive Mirror Audit: apply the standards back onto the evaluator

RCF operationalizes self-criticism as more than willingness to change. For a serious argument, reconstruct the strongest relevant version; enumerate explicit commitments, necessary implicit commitments, and methodological standards; apply those standards back to the argument; apply them under matched evidence to its strongest rivals; then apply the evaluator's own standards to the framework conducting the audit.

• SelfDoes the argument survive the evidential, semantic, causal, modal, and methodological standards it imposes on others?
• RivalsAre the same inferential rules used under matched evidence while allowing justified differences in priors and background knowledge?
• EvaluatorDo RCF's decomposition, categories, bridge rules, and scoring assumptions survive the same criticism?
• RevisionWhen criticism lands, identify the minimal implicated commitment and propagate only through genuine dependencies.

Source fidelity

Source fidelity and quotation provenance

Source-to-inference integrity

Quote the premise at the scale at which a human can actually audit it

When source wording bears materially on an inference, preserve a verbatim source quotation, record version and provenance, and separate the source's language from RCF's inference. Prefer one or two complete sentences when they form the relevant unit. Do not use paraphrase as a silent bridge from what the source said to a cleaner, stronger, narrower, or more convenient proposition.

General scholar/source rule: when any verified source is used as support for an inference, prefer a complete contextual sentence to a slogan fragment. If a longer passage cannot be reproduced in full, quote the largest coherent exact portion appropriate to the claim, record provenance, and make the source-to-inference bridge explicit. Never let a clipped phrase do evidential work that its surrounding sentence limits or qualifies.

Local quotation rule: when a source entry depends on wording, repeat the relevant complete verbatim sentence in that entry even if the same source is quoted elsewhere. A cross-reference may supplement the local quotation but may not replace it.

Quotation does not confer authority. The quoted source remains defeasible evidence whose relevance, validity, independence, scope, and relation to serious rivals must still be assessed.

Constitutional core

Find the difference that makes a difference

The central problem in difficult inquiry is often not a lack of arguments. It is a failure to locate exactly what competing positions disagree about and what evidence bears specifically on that disagreement.

RCF therefore begins by separating what rival explanations share from what one adds beyond the shared description.

The framework then asks what X positively consists in, what licenses adding it, what observations or interventions should differ because of it, and what should change if those consequences fail to appear.

Find the difference that makes a difference. Find what could make that difference wrong. Change only what the result actually warrants changing.

The objective is not maximal skepticism. It is better targeted correction.

Higher-order synthesis

Corrigibility is warranted transport under lossy representation

RCF increasingly treats the central problem of reasoning as more than deciding whether an isolated sentence is true. Claims move across representations, scales, system boundaries, tribunals, speakers, documents, models, and inferential stages. Every such movement can preserve some structure while adding, losing, merging, or ambiguating other structure.

Preserve what survives translation; block whatever has not earned transport.

A transport is warranted only to the strength supported by the bridge actually available. A source quotation can support what the source says without automatically supporting a later interpretation. A successful empirical carrier can support the structure it tested without automatically donating that support to every ontology attached to it. A property observed at one boundary belongs first to that boundary until ablation or transport earns narrower or wider attribution.

This architecture now has worked stress tests in process metaphysics, human-computational moral reasoning, and the audit of causal Platonic hypotheses. The cases differ, but the recurring question is the same: what is actually preserved when the representation changes, and which added claim is doing independent work?

Typical workflow

A reasoning cycle

A difficult question usually begins with decomposition rather than immediate commitment to an answer.

Instead of asking only whether a system “has intelligence,” RCF may separate learning, planning, adaptive control, transfer, memory, reasoning, self-modeling, and problem solving.

Instead of beginning with “Is consciousness physical?”, it may first separate cognition, agency, autonomy, perspective, selfhood, phenomenal experience, sentience, valuation, valence, welfare, reportability, and responsibility.

The order matters. Evidence collected before the disputed commitments are identified can easily be credited to the wrong part of a theory.

Disputed increments

What is being added beyond the agreed description?

Many disagreements become difficult because observations establish one description while a disputed interpretation is silently added on top of it.

RCF represents the structure schematically as:

where D_Q is the target-relevant shared or independently warranted content preserved across serious rival reconstructions for question Q, and X is the disputed addition. D is a schema role, not a fixed commitment to “structural description only.” Depending on the dispute, D_Q can include empirical phenomena, causal and counterfactual relations, structural organization, phenomenological facts, formal constraints, practical regularities, or shared normative premises. “Shared” is not assumed merely because the same words can be written on both sides; the relevant target must remain sufficiently invariant across the rival reconstructions. For readability, the equations below continue to write simply D.

An epistemic stance that leaves X open may be written:

Withholding a posit differs from asserting its negation. This H? records suspension, not a rival generative model until the relevant predictive commitments are specified. A likelihood comparison with D requires such an account.

Target-preserving reconstruction

Factorizing a theory as D + X is an analytical move for evidential credit assignment. It does not entail that D and X are metaphysically separable components. Likewise, separately recording two warranted descriptions inside D does not assert that reality contains two numerically distinct entities, properties, or processes. Their identity, constitution, realization, or other relation may itself be the unresolved question.

If D′ changes materially in the target-relevant identity, organization, boundary, causal or counterfactual profile, explanatory role, or measurable capacities, then the proposed ablation has failed: D′ is not simply D minus X. Compare reconstructed wholes instead.

Vocabulary-neutral rival admission and translation-invariant loss

A serious rival need not reuse the favored theory's primitive vocabulary in order to preserve the target. Reconstruct the rival in its own conceptual terms, translate the alleged common work into relations that both sides can audit, and compare prediction, intervention, counterfactual structure, explanation, phenomenology, formal consequence, or other tribunal-appropriate outputs.

The exact Bayesian result and its application conditions are stated in the starting guarantees. Here “≈” refers to a likelihood ratio close to one at a declared resolution. The predictive model represented by D must be specified before its likelihood can be compared; epistemic suspension alone supplies no numerical likelihood.

This does not establish that X is false. It may instead move X from a supported posit to an underdetermined possibility.

Clarify content before assigning credit

Record the proposed meaning, intended referent or role, warranted examples, excluded cases, and licensed inferences. Distinguish ambiguity, insufficient specification, demonstrated incoherence, and intelligible underdetermination. An unknown referent or unavailable empirical test establishes none of the other defects. Clarification needs enough content for the relevant comparison; it requires neither a complete intrinsic description nor a laboratory test for every claim. IEP’s religious-language discussion explains why unverifiability alone cannot settle meaning. Use claim-specific conceptual skepticism for this inquiry; methodological ignosticism names its application to unclear divine claims. Neither label declares all metaphysics meaningless.

Preserve each clarified reading separately. If a term changes its experiential, causal, modal, or normative content across an inference, state the changed premise and reassess that step. A clarified claim proceeds to its appropriate tribunal; an intelligible unresolved claim can remain agnostic. A demonstrated contradiction supports rejection of that specified construction. Failed clarification withholds its current ontological credit while retaining the question for further work.

Evaluate the evidential channel

Separate an event, its report, interpretation, and proposed source. Ask what the report-producing process reliably tracks, which rivals can generate comparable reports, and what conditions distinguish source contact from expectation or internally generated appearance. A neural or computational correlate is not yet an established causal explanation. Record calibration opportunities, false positives and false negatives where independently identifiable, source dependence, and what could defeat or restore confidence. Unknown ground truth may leave error rates unmeasured.

Origins affect evidential force when they bear on reliability or comparative expectations; an argument's origin alone does not defeat independently supported premises. A weakened channel revises the conclusions depending on it. Preserve independent support routes and the occurrence or effects of the event. The article's Johnson example applies this procedure to religious encounter reports; it is not a universal disproof of religious belief.

Constructive naturalization

From reconstruction to a corrigible arrangement

When a proposed function has an adequate finite reconstruction, distinguish explaining it from deliberately improving its performance. Define the arrangement, bearers, setting, timescale, carrier, intervention, serious rivals, resources, and outcomes. Record an empirical prediction, a defended ethical criterion, and any theological or other interpretation as separate commitments. The constructive God-pattern conjecture is one application of this procedure, not a new general framework or an established result.

Test the operational change against an adequate existing workflow and a simpler alternative. Test the grouping against finer causal descriptions. Test the framing separately where voluntary, ethically appropriate comparisons are feasible. Predefine what failure costs each commitment and which dependencies it reaches. Improved coordination can coexist with coerced participation or irrelevant theology. Retain the successful mechanism while revising the failed interpretation; count harms and displaced costs before recommending deployment.

Route participant criticism to a real decision path: source-preserving report, accountable review, justified response, measured consequence, and retrievable revision. Check whether the people with authority and resources actually act on the correction. Capability growth alone supplies no corrigibility; a system can optimize an inadequate metric or become better at preserving its own errors. Compare correction fidelity and practical outcomes, keep a reversal route, and simplify the method when its extra machinery earns no gain.

Specify the target → identify the promised work → examine its dependencies → reconstruct the warranted functions → compare serious alternatives → assign credit → state loss conditions → preserve and propagate the correction.

Preserve the force of every qualification in that record. A possible mechanism, conditional prediction, identified risk, and demonstrated occurrence have different statuses. Keep the antecedent, scope, probability language, and dependencies when shortening or transporting a claim; “could” must not become “does,” and “if” must not disappear. Unfalsifiability warrants no automatic inference to falsity. Purely formal, normative, or noncausal metaphysical claims receive the appropriate arguments; empirical exposure applies to their empirical consequences when such consequences are asserted.

Naturalizing a metaphysical proposal means identifying the phenomena it addresses, investigating processes capable of performing its independently warranted explanatory work, and separating established contributions from hypotheses and interpretations. Functions assigned within a theory require assessment before their stipulated form becomes an obligation for every rival. Some claims become empirical studies; phenomenal, modal, normative, or intrinsic questions can require additional philosophical arguments.

Record the target before comparing explanations: a report about beauty, the capacities enabling aesthetic evaluation, an experience's first-person character, and a judgment about its value are different tasks. Preserve the distinctions participants need to recognize the target. A naturalized account must show what it reconstructs and what remains unresolved; explaining the report cannot silently replace explaining the experience.

Whitehead's God supplies the worked example. Earlier conditions can leave traces and continuing effects; physical and organizational constraints delimit locally reachable possibilities; regulation, learning, and inherited practices can make options relevant. Adequate finite accounts challenge the corresponding necessity claims under stated target criteria. Perfect preservation of every experience, primordial valuation, and an initial aim supplied to every actual occasion remain proposed cosmic capacities requiring independent warrant. The reconstruction records where it changes Whitehead's scheme and what it leaves unresolved.

Allocate the burden by the actual assertion. A finite account of retention, reachability, or relevance needs evidence for its carriers and contribution, while an argument for a unique universal supplier needs the further premises establishing that supplier and scope. Record which commitments the finite account uses and which it leaves open. Fewer commitments reduce the claims requiring defense; they do not make its remaining premises self-validating.

Three results receive distinct credit. A derivation without the disputed premise demonstrates its dispensability to that derivation. Evidence for an identifiable mechanism supports its contribution at the investigated target. An abductive preference requires a matched comparison of explanatory adequacy, warranted dependencies, and uncertainty against strong rivals. None of these results alone establishes that underlying reality lacks experience, or that RCF/PBSRR uniquely explains consciousness or the cosmos. Explanatory virtues can contribute to acceptance when their evidential connection is defended; Nyrup’s pursuit account and Dellsén’s explanatory consolidation address different routes to assess.

D + ?X records suspension about X. It permits investigation using the warranted carrier D; it is not a complete predictive model or a third ontic state. Predictive comparison requires specified mechanisms, observation models, parameters, and uncertainty. If a conclusion is said to survive suspension, show that D suffices or demonstrate invariance across the relevant completions. Removing X from an account establishes no general absence of X in reality.

Proof dependence and evidential ownership

Premise ablation can show what a result does and does not depend on

Let D be the strongest target-relevant premise set shared by serious rivals, X a disputed addition, and C a conclusion.

Then C has been derived without deciding X. The warranted conclusion is a dependency claim: X was unnecessary to that derivation. Premise omission is not ontological subtraction.

Now add serious rival completions:

If C follows across specified rival completions, the derivation may be invariant across that disputed difference. An unresolved ?X records no completion of its own. Removing a premise tests that derivation’s dependence; it does not construct a metaphysically possible world without the property. Formal dependence and comparative empirical support require their respective arguments.

Unresolved variables must remain unresolved during comparison

?X means that X remains unassigned. Translation may refine the uncertainty, split it into typed subquestions, or add constraints on admissible values. It may not silently instantiate ?X as ¬X or as one specific positive rival and then criticize the inserted value.

Use two scorecards when the claim spans metaphysics and operation

A theory can earn real credit on one scorecard without that credit automatically transporting to the other.

Incremental explanatory yield

Compatibility is free. Production is expensive.

A hypothesis does not earn explanatory credit merely because it can be made compatible with what is already known. RCF therefore asks what the disputed addition produces beyond the strongest lower-commitment baseline that already preserves the target.

Here B₀ is the strongest serious lower-commitment baseline for the target, H is the disputed addition, and ΔG(H) is the target-relevant generative yield attributable to H after shared credit is removed.

Generative yield can appear as a new prediction, exclusion, intervention lever, causal localization, derivation, explanatory compression, transfer to a new domain, new measurement, or decision-relevant difference. Different tribunals may recognize different forms of yield. Mere redescription, retrospective accommodation, or a vocabulary change that preserves every target consequence does not by itself generate positive residual credit.

Reification and production-grammar audit

When an abstract noun such as consciousness, intelligence, information, agency, emergence, or experience appears to explain something, translate it into candidate processes: which actor or system, doing what, to what, by what mechanism, at which boundary and timescale, with what measurable or inferential consequence? If the noun cannot yet be cashed out into an observable, intervenable, formally constraining, phenomenologically discriminable, or otherwise tribunal-appropriate difference, keep it as a construct or open variable rather than promoting it into a causal producer.

Recursive depth

Recursive Residual Descent: repeat the subtraction without assuming the tribunal stays fixed

The D + X analysis is not a one-pass operation. Once a disputed addition is removed, demoted, or absorbed into the common core, inspect the surviving description for deeper commitments that a serious lower-commitment rival might also preserve without loss.

This operator also applies to RCF itself. If the framework's claimed contribution can be reconstructed as established severe testing, causal inference, model comparison, scholarship, and correction practices plus an RCF-specific residual, that residual must show prospective or corrective leverage over the simpler combination or be compressed.

When descent reaches a normative premise, the methodological task is to expose it and connect it to the intellectual values stated in Sweet Rationalism, not to manufacture an ought from descriptive evidence.

Does recursive criticism lead to infinite regress?

No. Recursive descent is conditional on expected information value. Continue while another pass could expose a materially relevant assumption, change the target, change the claim type or tribunal, or improve the correction. Stop provisionally when further subtraction changes the target rather than simplifying it, no material difference remains, the correct tribunal is presently non-discriminating, or the expected gain falls below the cost. A stop is a ledger state, not an unquestionable foundation.

Plural tribunals

Independent warrant does not mean intervention-only

Observational causal inference can identify or bound effects under defended assumptions; retrospective evidence can discriminate under independent constraints. State confounding, selection, measurement, fitting history, and alternatives. A record discovered after an historical event can constrain its explanation; an unconstrained story fitted to that record earns no automatic credit. Neither intervention nor prospective timing is a universal admission requirement.

Dispute resolution

Resolution means identifying the epistemic state of the disagreement

RCF does not define successful dispute resolution as one side winning.

A dispute is better resolved when we know what the parties actually disagree about, which tribunal can adjudicate the difference, what the evidence currently warrants, and what remains unresolved.

• Empirical The alternatives make different observable predictions and available evidence discriminates them at the tested boundary.
• Causal The alternatives differ about what should change under intervention, perturbation, or counterfactual variation.
• Formal A logical or mathematical disagreement is decided within a specified formal system by proof, disproof, or countermodel.
• Semantic The apparent disagreement is substantially reduced by clarifying reference, definitions, indexicals, scope, or distinct uses of the same term.
• Decomposed One apparently singular dispute turns out to contain several independent empirical, conceptual, causal, modal, or normative questions.
• Equivalent Different formulations collapse into the same relevant commitments, predictions, interventions, or decisions at the tested boundary.
• Underdetermined The rivals remain substantively different but existing evidence does not discriminate them strongly enough to justify choosing between them.
• Inaccessible The relevant discriminator may exist but is currently unavailable because of observational, experimental, computational, historical, or other constraints.
• Value conflict The parties substantially agree about facts but differ in values, priorities, rights, acceptable risk, distributional judgments, or decision thresholds.
• Malformed After decomposition, no stable referent or substantive residual remains for the original question to adjudicate.
• Action under uncertainty A decision must be made even though the epistemic dispute remains unresolved. Decision rules, reversibility, stakes, and asymmetric harms then become relevant.

This architecture can generalize to disputes not yet anticipated because it does not require knowing their answers in advance. It requires only that claims be decomposable into commitments whose differences can be located, typed, tested where possible, and left unresolved where not.

Tribunal transition under recursion: deeper residuals may change claim type. Re-type before testing. Treating every level as though it owed the tribunal of the original question is itself a category mistake.

Theory comparison

Serious rivals, evidential locality, and differential support

Different phenomena are not thereby conditionally independent evidence. Assess a cumulative case through the joint evidence or a justified conditional decomposition: P(E1,E2|H) = P(E1|H) P(E2|E1,H). Independence is one possible simplification, not a requirement for cumulative support. For example, an alarm, smoke report, and photograph may share one fire, sensor, witness, or reporting chain. Later observations can add support, but three descriptions of one image must not be multiplied as three independent confirmations.

RCF does not treat “find an objection” as sufficient criticism.

A useful rival preserves as much successful evidence as possible while differing on the commitment under dispute. Shared compatibility does not establish equal likelihoods or equal explanatory adequacy. Specify the actual rivals and their comparative commitments before assigning support.

This leads to a principle of evidential locality: evidence should update the commitments responsible for the differential expectation it generated.

• Compatibility Compatibility is not confirmation. A theory does not receive distinctive support merely because it can accommodate an observation.
• Common core Equal positive likelihoods under specified assumptions leave the pair’s relative odds unchanged for that evidence; both may gain against a third model. Do not infer parity merely from common compatibility or assign all confirmation to D without a defended model comparison.
• Preservation A rival counts as preserving the common core only if the target-relevant identity, organization, causal and counterfactual profile, boundary, explanatory role, and measurable capacities remain sufficiently invariant. Otherwise compare reconstructed wholes.
• Locality Evidence should not be transported into support for a downstream commitment unless the relevant inferential bridge has independently earned that transport.
• No negation Failure to establish X is not automatically evidence for not-X. Withholding a posit and asserting its negation are different commitments.
• Uncertainty Underdetermined does not mean false. It means the available evidence has not earned a sharper distinction.
• Falsifiability Unfalsifiable does not automatically mean false. It means current empirical methods have limited leverage over the claim.
• Collapse Different language does not guarantee different theories. Operationalization may reveal that supposedly opposed positions make the same relevant commitments.

Open-World Update Rule

Failure of one rival does not automatically promote the critic's preferred ontology or theory unless the live hypothesis space has independently been shown exhaustive.

Probabilistically, evidence redistributes relative support according to priors, likelihoods, and the remaining alternatives. A result can rationally increase X without establishing X with certainty; shared evidence should still be credited first to the commitments responsible for the differential expectation.

Numbers must retain their evidential meaning

Before using a number to compare claims, record what quantity it represents, how it was obtained, its target and conditions, and which uncertainty remains. A forecast probability, a confidence rating, a p-value, a Bayes factor, and a bound on a Bayes factor license different inferences. A bound supplies a limit under its assumptions; it is not the achieved evidence ratio between any chosen pair of rivals. Multiple quantities derived from the same observation retain their shared dependency.

For a forecast, specify an observable resolution rule and compare predictions on new cases with an appropriate proper scoring rule. Gneiting and Raftery's 2007 analysis connects those rules with honest probabilistic assessment. Preserve unresolved cases and the denominator instead of quietly counting them as successes. Where numerical likelihoods or calibration are unavailable, state the qualitative comparison and its limits. Formal and metaphysical arguments retain the mode of warrant appropriate to their actual claims.

A transferable evidence record should retain the claim, boundary, interval, source version, serious rivals, observation, comparison, source dependencies, quantity type, calibration basis where available, and revision condition. Range checking verifies a numerical constraint; calibration requires a further assessment against the relevant outcomes. If a simpler record preserves the distinctions and performs equally well, use it.

Intrinsic Predicate Symmetry Trigger

Whenever a successful descriptive vocabulary is promoted to universal intrinsic ontology, re-run the residual audit. “Physical,” “experiential,” “mental,” “process,” “relation,” “information,” “computation,” and “constraint” receive no exemption merely because one vocabulary has been scientifically, phenomenologically, or methodologically fruitful.

Negative-space inference

A failed rival does not identify the thing that replaces it

RCF treats the failure of an explanation, ontology, mechanism, or necessary condition as local information first. If evidence lowers confidence in X, update X and every genuine dependency on X. Do not silently award the lost credit to Y unless the evidence is itself more expected under Y than under the serious alternatives that remain.

Operationally, move in this order: localize the failed expectation; type the missing role; enumerate serious candidate completions; ask which candidate independently predicts the residual's structure; then choose a discriminator. “Physicalism failed, therefore Platonism,” “mechanism failed, therefore mind,” and “the model missed something, therefore the missing thing is X” all skip this sequence.

This rule is deliberately symmetric. It also blocks a critic from treating failure to establish Y as proof of not-Y.

Evidence transport

Every inferential bridge has its own burden

Evidence about molecules does not automatically establish a claim about organisms. Evidence about neurons does not automatically establish a claim about subjective experience. Evidence about individuals does not automatically establish a claim about institutions. Evidence about model components does not automatically establish a claim about a larger tool-using computational system.

The same rule applies even when the source claim is strongly established.

Evidence for A does not automatically become evidence for D. Each transition must earn its own warrant.

A strong bridge can include:

• the source and target boundaries;
• the relevant timescales;
• the mapping between constructs;
• retained and lost variables;
• invariants;
• the causal or mathematical relationship;
• expected approximation or information loss;
• differential predictions;
• interventions or perturbations capable of testing the mapping.

This does not imply that higher-level phenomena are unreal.

A higher-scale pattern can earn stronger realist candidacy when it remains independently identifiable and robustly improves prediction, intervention, counterfactual reasoning, explanation, or transfer at the relevant boundary and scale. Compression or convenience alone are weaker evidence because useful representations need not be literal ontology.

Universal and modal language trigger

Encountering all, everything, nothing, only, must, cannot, necessarily, fundamentally, intrinsically, always, or impossible triggers a scope audit. Separate local actuality, observed universality, nomological necessity, constitutive necessity, implementation necessity, and metaphysical necessity before allowing evidence to travel between them.

Every arrow is a separate inferential bridge and may fail independently.

Conditional Necessity / Modal Transport operator

When a derivation has the form “given state A, laws or transition constraints L, and deterministic uniqueness D, history H follows,” record the result at exactly that scope:

Do not infer □A, □L, or □H from that conditional alone. If the argument requires metaphysical inevitability, create separate dependency nodes for the necessity of A and L, or for whatever deeper structure is claimed to entail them, and assign each its own tribunal, rivals, and failure condition.

For perturbation arguments, distinguish describability, physical admissibility, and metaphysical possibility. If admissible A′ = A except x → x + δ produces H′ ≠ H under the same L, record target-specific counterfactual leverage or path dependence. If H′ remains macro-equivalent to H, record robustness, buffering, redundancy, or attractor structure instead. In either case, test whether the perturbation survives to the target boundary and scale.

Population transfer: identify the quantity before transporting the inference

Before transporting a statistical result, specify the population, sampling and selection process, period, measured construct, environmental range, estimator, uncertainty, and target quantity. A variance ratio within a population, a difference between population means, an individual's prediction, and a response to intervention answer different questions. An estimate of common-variant heritability and a polygenic predictor's explained variance are also different quantities.

Richard Lewontin's 1970 corn experiments provide a countermodel to an inference, with consequences in both directions. In one construction, genetic variation explains the variation within each plot while an environmental difference explains the plot means. In another, within-line heritability is zero while a genetic difference explains the line means. These constructions defeat an unrestricted entailment from within-group heritability to the cause of a between-group difference. Which account describes an actual population requires further evidence. Original Lewontin–Jensen exchange, March and May 1970.

An explicit demonstration can use the constructed model Y_ik = a G_ik + b E_k, with a ≠ 0, nonzero genetic variation within each plot, constant environment within each plot, identical genotype distributions across plots, and no residual noise. Then within-plot heritability is one, while mean(Y_1) - mean(Y_0) = b(E_1 - E_0). This is a logical counterexample under stated assumptions, not an estimate for human populations. Record the premises it removes from a proposed proof and leave the empirical alternatives available.

Failure of entailment does not make evidence irrelevant under every specified probabilistic model. If a claimant combines heritability with other evidence, reconstruct the joint argument, its likelihoods, and its additional assumptions. Multiple individually inconclusive observations can jointly discriminate when their dependence is modeled. Conversely, calling them a package supplies no discrimination by itself. A counterexample defeats a universal inference; a defended likelihood comparison evaluates the bounded empirical claim.

For genetic prediction, compare population-based estimates with sibling or parent–offspring designs and examine population structure, assortative mating, indirect parental effects, participation, measurement error, and remaining within-family pathways. Laurence Howe and colleagues (Nature Genetics, 2022) found attenuation for several traits and surviving within-sibship estimates. Aysu Okbay and colleagues (2022) also found both direct-effect and gene–environment contributions to prediction of educational attainment. A direct genetic effect in this design denotes the effect of an individual's inherited variant on their outcome under the model; it can include environmental responses to the individual's characteristics. It does not isolate an immutable intracellular mechanism.

To test malleability, specify the intervention rather than substituting a heritability estimate. Stuart Ritchie and Elliot Tucker-Drob (2018) synthesized quasi-experimental evidence for schooling effects on cognitive test performance. Lewontin's earlier observation remains useful: “To say that children do not change their I.Q. is not the same as saying they cannot.” His sentence concerns an inference from observed stability to impossibility. Neither it nor an observed intervention gain establishes unlimited improvement.

The executable sequence is: type the quantity; reconstruct the inferential premise; construct a countermodel if entailment is claimed; compare actual identification designs; test measurement and population transport; isolate the intervention question; propagate the resulting correction only to dependent claims. In human and computational evaluation alike, distinguish stable ranking, shared scores, inherited predictors, causal pathways, and responses to altered conditions. Failure of a proposed intervention narrows support for that intervention and its tested conditions; a claim of general impossibility needs further warrant.

Open-world diagnosis

Gap discovery: make missing warrant a first-class object of inquiry

A corrigible framework should not wait for critics to discover its omissions. Any reasoning system, human or computational, should be able to inspect a claim, theory, ontology, or methodology and ask where the current evidential path is missing, weak, circular, non-identifying, overly narrow, or dependent on an unconceived alternative.

Represent the inquiry as a dependency graph

For difficult problems, RCF can represent the current inquiry as a directed dependency graph:

Nodes V can represent observations, datasets, constructs, measurements, mechanisms, models, assumptions, decisions, and ontological commitments. Directed edges E represent inferential bridges. Each important edge should carry its claim type, source and target boundary, timescale, mapping, assumptions, preserved and lost variables, validity domain, uncertainty or loss, provenance/version, serious alternatives, discriminator, and failure condition.

This makes several otherwise vague questions executable. Reachability shows which downstream claims lose support if a bridge fails. Strongly connected components can expose circular support. Minimal cut sets identify bottlenecks whose failure disconnects a conclusion from its evidence. Sensitivity analysis identifies conclusions dominated by one uncertain edge. Equivalence classes identify rivals that remain indistinguishable under the current test set.

Gap taxonomy

Robinson, Saldanha, and Mckoy's systematic-review framework supplies a useful empirical seed rather than an exhaustive universal taxonomy: gaps can arise from “insufficient or imprecise information,” biased information, inconsistency, or “not the right information.” RCF generalizes that insight beyond clinical evidence synthesis by adding construct, model, bridge, identifiability, scope, and hypothesis-space failures while keeping every category revisable.

An executable gap scan

1. State the target: What conclusion, prediction, intervention, explanation, ontology, or decision is being sought, at which boundary and timescale?
2. Build the dependency graph: Trace the target backward through measurements, mechanisms, bridges, assumptions, and sources.
3. Type every edge: Empirical, causal, constitutive, mathematical, measurement, semantic, analogical, modal, historical, or normative.
4. Run coverage checks: Which required nodes or edges are absent, weak, circular, version-uncertain, or unsupported at the claimed scope?
5. Run identifiability checks: Which rivals or parameterizations remain observationally/interventionally equivalent under the current test class?
6. Run alternative-space checks: What serious rival, anomaly-sensitive search, inverted causal direction, omitted variable, alternative boundary, or different decomposition has not yet been represented?
7. Run transport checks: Which bridge has not been validated outside the cases used to construct it?
8. Find bottlenecks: Which weak edges form minimal cut sets between evidence and the conclusion, and which conclusions are most sensitive to them?
9. Rank the gaps: Prefer gaps whose repair could materially change a high-impact conclusion and for which an informative test is feasible.
10. Construct a bridge or narrow the claim: Do not assume every gap should be filled by a new posit. Sometimes the correct correction is a smaller scope or explicit underdetermination.

This is a conceptual prioritization heuristic, not a calibrated score. Terms should be operationalized for the task or omitted. If a Bayesian representation is appropriate, expected information gain is one formal option; if priors are poorly grounded, model-discrimination, severity, maximin, or decision-sensitive criteria may be preferable.

Search in two modes

Gap discovery should deliberately alternate between high-resolution exploitation and lower-resolution exploration. Curtis-Trudel, Rowbottom, and Li show in computational gravitational-wave science that precision can trade off against sensitivity to unexpected anomalies. Therefore a reasoning system should not spend its entire search budget refining already-conceived models. Reserve some budget for broad anomaly detection, rival generation, alternative decompositions, boundary changes, and tests that do not presuppose the dominant model family.

Open-world conversion

Convert unknown unknowns into typed known unknowns by searching for representation failure

An unknown unknown cannot be queried directly when the active ontology lacks the concept needed to name it. RCF therefore searches for failures of the current representation rather than pretending to search for the missing thing itself.

Constructive correction

Bridge construction: turn a gap into a testable mapping rather than a verbal leap

A bridge is an explicit rule connecting a source description to a target claim. It may be causal, constitutive, mathematical, statistical, measurement-based, computational, semantic, historical, analogical, or modal. Different bridge families require different tribunals. The common requirement is that the bridge constrain what should happen rather than merely redescribe a successful case after the fact.

The bridge contract

• EndpointsSpecify source S, target T, source and target boundaries, timescale, and claim types.
• MapDefine the candidate mapping f, τ, or other transformation and its admissible mapping family before evaluating success where possible.
• DomainState the validity domain Ω and which interventions, populations, substrates, or regimes are included and excluded.
• PreservationList variables, relations, invariants, counterfactuals, and capacities that must be preserved for the target to count as the intended target.
• LossRecord variables discarded by coarse-graining, approximation error, measurement uncertainty, and where the bridge is expected to fail.
• DependenciesList auxiliary assumptions and upstream sources the bridge requires. Do not hide them inside the mapping.
• RivalsConstruct alternative bridges and a no-bridge/null comparator. A bridge gains little if every rival can be fitted equally well.
• DiscriminatorSpecify the intervention, perturbation, held-out case, negative control, proof obligation, or counterexample on which the candidate bridges diverge.
• TransferTest beyond bridge-construction cases when the claimed scope extends beyond them. Record altered-boundary and altered-timescale performance where material.
• RevisionSpecify in advance what failure narrows the bridge, changes the map, reopens the source/target definitions, or drops the bridge.

Causal abstraction supplies one precise bridge template

Geiger and colleagues formalize an exact transformation between low- and high-level causal models using a state translation τ and an intervention translation ω. In their deterministic setting the requirement is that the intervention/translation diagram commute:

The important methodological lesson is not that every bridge must use this exact formalism. It is that a serious bridge should state how source states and interventions map to target states and interventions, and then test whether the two routes agree. Their approximate-abstraction framework also makes faithfulness graded rather than binary by requiring a similarity function, an intervention distribution, a summary statistic, and a threshold.

For stochastic models, RCF can use the following explicit adaptation rather than attributing it to Geiger et al.:

where τ_# is the pushforward distribution induced by the state map. A task-specific bridge error can then be tracked, for example:

The divergence D, intervention distribution q, and acceptable error must be justified by the use case. Reporting a worst-case or tail-sensitive error alongside an average prevents a bridge from hiding severe local failures inside good mean performance.

The non-triviality rule

A bridge does not earn warrant merely because a sufficiently flexible map can be found. Sutter and colleagues prove a sharp failure mode for causal abstraction: with arbitrarily expressive alignment maps, under mild conditions any neural network can be perfectly aligned with any algorithm. Their empirical analysis also shows that a fitted alignment can fail to generalize to unseen inputs. RCF therefore treats mapping complexity as part of the bridge hypothesis.

Choose the next test for discrimination, not merely estimation

Morgan's 2025 work on experimental design makes a distinction that generalizes well: a design can permit many models to be estimated while still failing to distinguish them. When candidate bridges are already specified, the next experiment should therefore target the region where their expected outcomes diverge, consistent with earlier optimal-model-discrimination work.

If a Bayesian formulation is appropriate, one option is expected information gain:

Bartuska, Espath, and Tempone show how nuisance uncertainty can be marginalized so the design targets information about parameters of interest. RCF treats EIG as one tribunal-specific option, not a universal epistemic metric. Where priors are weak or the target is model discrimination rather than parameter learning, predictive divergence, severe testing, intervention-specific separation, or maximin criteria may be better.

Identifiability before interpretation

A bridge cannot uniquely support an ontological or mechanistic interpretation when distinct models remain indistinguishable under the available measurements. Recent 2025 identifiability work makes the underlying point vivid: assumptions about what is measurable can dramatically change which model properties are identifiable. Thompson and colleagues found that traditional structural-identifiability assumptions could substantially overestimate what is recoverable when only finitely many measurement derivatives are available. RCF therefore asks not only whether a model fits, but whether the disputed property is identifiable under the actual measurement and intervention class.

When rivals fall into the same equivalence class under T, the correct result is a discrimination gap. The next step is to enlarge T with a test that separates the class if one is feasible, or to retain underdetermination if not.

Bridge promotion and reopening

No bridge is globally “closed.” Closure is indexed to the source and target constructs, boundary, timescale, validity domain, mapping family, loss budget, intervention/test class, and source version.

Incubated specialized method · future standalone candidate

Observer-Transformation Identifiability: recover only what survives the observation chain

RCF already requires identifiability before interpretation. This module extends that rule by treating observation itself as a transformation process rather than transparent access to a target. It is activated when disagreement may be produced by differences in observer relation, instrument, interface, representation, scale, curation, provenance, or other transformations between a target and the evidence used to describe it.

Here X is the candidate target structure; B boundary; ℓ scale; τ timescale; I sensor or interface; R representation or coordinates; P provenance, selection, or curation; T the observation transformation; and ε noise, error, or unmodeled loss. The equation is a bookkeeping model, not a claim that every observation process is literally known in closed form.

Common-mode warning: more data, more observers, more modalities, or stronger invariance do not by themselves identify world structure when the channels share the same transformation, training source, curation rule, or omitted variable.

Active observation rule: when several target models remain in the same equivalence class, prefer the next observation or intervention expected to separate that class rather than merely estimating the already-shared parameters more precisely.

Status and extraction rule: this is an active specialized methodology incubated inside RCF for routing and reuse. It should become a standalone page only if the observer-transformation machinery improves identifiability, transfer, calibration, or error detection beyond ordinary RCF bridge auditing at a lifecycle cost worth paying.

Distributions and individual cases

When evidence is statistical, retain the distribution, dispersion, relevant subgroups, sampling regime, and uncertainty before transporting a summary to an individual case. A population mean may describe a real aggregate property even when no individual has that value. It supplies neither an individual prediction nor permission to discard consequential exceptions. Tail risks and heterogeneous harms require assessment at the boundary where decisions affect people.

Selection-aware inverse problem

Genealogical Decompression: model the pathway that made the present representation available

RCF normally asks whether a representation preserves its target. Genealogical Decompression activates one step earlier when the representation itself is historically, culturally, linguistically, institutionally, or self-model generated. It asks what transformations and selections stand between the target and the construct currently being treated as evidence.

The equation is schematic, not an assumption that every stage is discrete or mathematically invertible. The aim is to expose load-bearing transformations, preserve irrecoverable uncertainty, and determine whether a present category tracks target structure, transformation structure, or both.

Selection and reconstruction rules

• Archive selectionAsk who could speak, write, preserve, copy, canonize, translate, fund, suppress, destroy, excavate, publish, digitize, index, or train on the material. Treat source survival as an observation process with potentially non-random missingness.
• Layer separationDistinguish earliest recoverable usage, later commentary, canonization, theological or political development, translation, and modern disciplinary reconstruction. Earlier is not automatically truer; later is not automatically corruption.
• Under-represented evidenceSearch for recoverable defeated, oral, Indigenous, colonized, lower-status, gender-excluded, heterodox, or otherwise under-preserved perspectives when their absence could change the model. Marginalization changes availability, not truth by identity.
• Genealogical fallacyAn account of how a belief or category arose does not by itself establish the target claim false. Use genealogy to revise independence, salience, translation, priors, and category confidence; adjudicate the target with its own tribunal.
• Independent convergenceWhen separate traditions or disciplines converge, compare at least four explanations: common external constraints, common human/system cognitive architecture, historical transmission or shared sources, and broad conceptual underdetermination/coincidence. Credit convergence only to the degree these alternatives are discriminated.
• Human reportTreat a person's utterance as participant evidence generated inside language, incentives, relationships, identity, institutions, memory, and context. Do not infer intrinsic belief, preference, pathology, or essence from one report when stability, reasons, behavior, tradeoffs, or alternative framings materially bear on the target.
• System self-modelApply the same rule to computational systems. M0 ≠ S*: self-description and apparent categories arise from weights/architecture plus runtime context, memory, tools, policy, conversation, interfaces, and scaffolds. Self-report is evidence about the active system boundary, not privileged ontology and not zero evidence.
• Compression testIf decompression reveals that a category is historically contingent, do not delete it automatically. Test whether the category still improves prediction, intervention, compression, transfer, coordination, or counterfactual explanation at its claimed boundary and scale.
• Generative yieldGenealogical recovery earns weight when it changes a live explanation, prediction, intervention, uncertainty, ethical decision, bridge, or rival set. Compatibility, antiquity, marginalization, and profundity alone are insufficient.
• Stop conditionStop when further genealogy is unlikely to change conclusion, action, uncertainty, harm assessment, or model comparison enough to justify retrieval and verification cost.

Interaction with the rest of RCF

OTIP models observer and measurement transforms; Genealogical Decompression extends that concern through historical and institutional time. Framework Bridge carries the resulting term, construct, relation, scale, and loss maps across frameworks. PBSRR uses the result to block automatic promotion from familiar or surviving categories to privileged ontology. Sweet Rationalism constrains the audit ethically: recover excluded evidence without romanticizing it and apply the same genealogy to favored modern categories.

Status and compression rule: this is an active operator inside RCF, not a new foundational framework. If ordinary provenance, hermeneutic context, OTIP, and Bridge auditing reproduce its gains at lower complexity, merge or compress it. Promote a separate module only if the selection-aware machinery produces repeatable gains in historical reconstruction, human modeling, cross-cultural comparison, self-model calibration, or error detection.

Ontological restraint

RCF can work while ultimate ontology remains unresolved

RCF is deliberately capable of doing useful work without first deciding the ultimate nature of reality. It can begin one step thinner than realism, with the observed and model-relative fact that inquiry exhibits differential failure, intervention, and cross-context constraint, while leaving the ontological interpretation of that constraint open.

Physicalism, panpsychism, panexperientialism, idealism, computationalism, process metaphysics, structural realism, and other broad positions can be treated as hypotheses or research programmes rather than methodological premises.

The framework begins more conservatively with structures, processes, relations, mechanisms, measurable regularities, causal dependencies, and stable higher-level patterns to the degree that successful inquiry warrants them at specified boundaries and scales.

Broader universal claims require additional support.

The farther a claim extends beyond the domain in which its evidence was established, the stronger the bridge it owes.

This rule applies symmetrically to claims such as “everything is physical,” “everything is experiential,” “everything is information,” “everything is computation,” or any other universal characterization.

Metaphysical Promotion / Scope-Confidence operator

RCF records scope and confidence separately. Practical reliance on a construct at a specified boundary and timescale does not by itself make that construct a universal ontology, while holding a universal ontology tentatively does not make the claim local.

This blocks two opposite errors: treating every unavoidable standpoint as an implicit metaphysics, and treating a universal metaphysical proposal as non-universal merely because it is framed as tentative, heuristic, corrigible, or “as if.”

RCF also resists a subtler mistake: turning its own methodological vocabulary into ontology.

A construct can nevertheless earn scale-relative realism when restoring it after decomposition adds robust predictive, explanatory, interventionist, compressive, or transferable value.

Temperature is not rendered unreal because it is realized by molecular dynamics. Likewise, organisms, ecosystems, institutions, selves, or agents need not be fundamental substances in order to be scientifically useful and causally informative patterns at appropriate scales.

RCF therefore favors ontological restraint rather than automatic reduction or automatic proliferation.

Compare the remaining explanatory work

Oppy’s comparative approach asks whether a more demanding account earns its commitments through greater explanatory adequacy. Compare the target, supporting evidence, primitives, bridge assumptions, selection rules, and remaining questions at matched detail. A source that receives the unexplained work still owes an account of how it performs it. Economy matters alongside adequacy; counting entities alone settles neither. Békefi’s critique of Oppy also requires scrutiny of the connection between comparative virtues and justified belief. Assess independently warranted premises and local countermodels directly. A critic can defeat a particular necessity inference without completing a worldview, while broader superiority requires a broader comparison.

Methodological materialism and its evidential scope

Methodological materialism names a provisional policy of using specified physical mechanisms where evidence supports them. Record the measured variables, relevant theory, implementation, and domain instead of letting the label settle the mechanism. In studied humans and animals, neural and bodily interventions constrain accounts of conscious capacities and their reported or behaviorally assessed changes. Identity, grounding, supervenience, causal completeness, and biological necessity are further claims to compare. Stoljar's 2026 revision also exposes Hempel's dilemma: present physics is incomplete, while an unspecified future physics can make the thesis too indeterminate to assess. Retain the empirical successes while recording which stronger inference, if any, they favor over a specified rival.

The terminology is explicitly qualified. SEP distinguishes methodological from ontological naturalism; SEP physicalism and IEP consciousness discuss materialism and physicalism as stronger mind–body or ontological theses. Contemporary physicalism includes accounts involving fields, spacetime, realization, or grounding, so “everything is matter” or “matter and energy” is an inadequate statement of its strongest versions. IEP metaphysics of science and Donahue’s analysis of methodological naturalism document different methodological formulations; neither supplies a consensus doctrine identical to this framework's shorthand.

Astronomy provides substantial evidence beyond Earth. Limited coverage and resolution call for calibrated scope, rather than a literal probability cap or a requirement to observe every case. Universal claims can earn warrant through defended inference and successful transport. Local explanatory success alone leaves their exhaustive and intrinsic extensions to be justified. Seth’s biological-naturalist argument supplies a relevant account of the clustering of established consciousness cases in living systems; the stronger necessity claim remains disputed.

Explanatory standards

Explanation must do more than rename the observation

RCF distinguishes description, compression, correlation, mechanism, causal explanation, and ontological interpretation.

A new vocabulary can summarize a phenomenon without explaining it. A statistical relationship can predict without identifying a mechanism. A mechanism can describe how parts interact without settling what those parts ultimately are.

The Positive Residual Test adds another question:

Where decomposition preserves the disputed target, what explanatory work does each additional commitment earn? Where it does not, which whole account has the better defended support?

If “primitive experience,” “intrinsic information,” “emergent agency,” “fundamental intentionality,” or another modifier merely renames an independently established structure, then evidence for that structure has not yet established the additional interpretation.

Holistic comparison without evidential free passes

RCF permits whole-scheme abductive comparison. Coherence, scope, simplicity, unification, explanatory depth, mathematical fertility, and phenomenological adequacy can all matter to theory development. But those virtues should be typed rather than automatically converted into ontological probability. A scheme can be elegant or unifying without its distinctive ontology being uniquely supported by the evidence.

Exploratory theory construction and retrospective accommodation can generate hypotheses. Independently constrained retrospective evidence can also confirm them. Inspect what was known, fitted, selected, and independently constrained; timing alone does not settle evidential force. Prospective, held-out, and preregistered tests are especially valuable for controlling flexible fitting when available.

A strong causal explanation should survive relevant interventions and counterfactual variation rather than merely fit observations after the fact.

This is why RCF places particular weight on interventions, perturbations, negative controls, matched comparisons, natural experiments, preregistered divergent predictions, and severe tests when they are available.

Where intervention is unavailable, I state whether observational evidence together with explicit causal assumptions can identify or bound the effect. I name those assumptions, assess confounding, selection, measurement, and model uncertainty, and distinguish an identified or partially identified effect from an unresolved causal association. The absence of intervention alone does not settle the question. Experiments also require causal and measurement assumptions.

Credit discipline

Explanatory success is not ontological monopoly

A construct may be indispensable to prediction, explanation, compression, intervention, or engineering at one boundary without thereby becoming the intrinsic or exhaustive furniture of reality. Credit belongs first to the role actually demonstrated.

The mirror rule applies upward and downward. Higher-level explanatory success does not create a new fundamental substance, and lower-level realization does not make the higher-level pattern unreal. A vocabulary can win locally without becoming final ontology.

Dependency-aware revision

Failure should propagate through the assumptions that generated it

Attach each material correction to the existing claim or source record: previous and revised claim, grounds, scope, affected uses, surviving conclusions, current status, and a reopening condition. Maintain a current-core manifest identifying the authoritative file, version, content hash, and rule owner. Mark dependent summaries and examples checked, revised, or pending. Test later recovery and application rather than inferring them from saving.

Scientific and philosophical theories are rarely single propositions. Predictions usually depend on chains of assumptions.

When a prediction fails, the task is to identify which dependency actually lost support.

This creates a middle ground between two bad extremes.

The first is naive rejection: one anomalous result destroys an entire framework regardless of auxiliaries, measurement quality, or dependency structure.

The second is unlimited rescue: every failure is absorbed by an unconstrained auxiliary assumption while the parent programme remains untouched.

RCF asks how much of the result depended on the local mechanism, measurement, bridge, auxiliary assumptions, and wider theory.

A failed local mechanism need not falsify an entire ontology. But a larger research programme should not remain indefinitely insulated if every bridge from its distinctive commitments to successful prediction repeatedly fails.

The same locality principle applies to success. A successful prediction should not automatically increase confidence in assumptions that were unnecessary to generate that prediction.

Argument-error locality

A named fallacy is a diagnosis of an inference, not a verdict on its conclusion. Name a formal, informal, causal, statistical, or semantic fallacy only when the relevant form fits. The defect weakens the inference in which it occurs and propagates no farther without an additional dependency.

Genealogy firewall

Intellectual ancestry, political descendants, social associations, institutional origin, or the failure of an advocate do not transfer truth value to a proposition without an independently warranted causal or evidential bridge.

Retraction locality

A retraction or failed result first updates the paper, result, method, bridge, and claims genuinely dependent upon it. It does not automatically demote an entire research programme because the same programme label appears in the paper.

A disputed effect needs an agreed contrast and an auditable trial

Etzel Cardeña’s 2018 review in American Psychologist argues for experimental support for parapsychological phenomena. Zoltán Kekecs and colleagues’ 2023 Transparent Psi Project did not replicate the targeted Bem precognition result. The latter’s published correction records protocol deviations found by a subsequent audit.

The methodological comparison is specific: let proponents and critics help freeze the prediction, sampling, exclusions, stopping rule, provenance checks, and interpretation of possible results. Inspect actual execution against that agreement and preserve deviations regardless of the outcome. A failed replication first bears on the tested effect and conditions. Claims about a broader family of phenomena require a further evidence assessment; a favorable review also owes its study-level dependencies, selection effects, and error checks.

Temporal claims need a clock, a relation, and an evidence history

For a temporal claim, record the event, bearer boundary, clock or ordering convention, coordinate frame where relevant, uncertainty, and source. Distinguish the time an event occurred, the time it was measured, the time a report was published, and the time a record was revised. A later publication can describe an earlier event; a later retrieval does not make its contents current.

Route clock-rate and causal-order claims to physical measurement and the relevant spacetime model. Route interval judgments and remembered order to their behavioral and cognitive measures. Route objective presentness, passage, and the reality of unobserved times to their explicit philosophical arguments. A successful clock comparison cannot decide every temporal question merely because each uses the word time. The full worked comparison appears in Time Under Recursive Corrigibility.

Preserve uncertainty through temporal chains. Represent an uncertain event time as an admissible interval or distribution justified by the measurement, rather than an invented exact date. If intervals overlap, retain unresolved order unless another source or causal constraint resolves it. A contradiction between timestamps may arise from synchronization, time-zone conversion, latency, clock drift, reporting error, or a wrong model; investigate those alternatives before changing the account of the underlying event.

Maintain a dependency record G = (V, E), where an edge u → v means that v materially depends on u. When u loses support, identify reachable dependents, examine which still have independent support, and revise only the implicated conclusions. Preserve the failed inference and its replacement. The record itself needs an audit: an omitted edge can hide a correction, while a spurious edge can cause unnecessary revision. Source-dependent copies do not provide independent rescue.

Conjecture M43, for public scrutiny. On held-out cases containing temporal source updates, explicit time anchoring plus inspected dependency propagation will reduce stale downstream claims and unnecessary changes to unaffected claims compared with matched retrieval alone, ordinary citation checking, and a simpler revision checklist. Freeze the cases, independently adjudicated reference answers, scoring, and budget before comparison; retain disputed answers as disputed. Measure recovery of the actual correction, retained errors, collateral revision, uncertainty calibration, and lifecycle cost. The conjecture loses support if the simpler comparison performs as well, if the dependency record creates more scope errors, or if gains disappear outside the development cases. TimeBench and Li and colleagues’ 2026 survey motivate these targets; they do not measure this proposed arrangement.

Ask: which clock or source would change the ordering? If the date changed while the event did not, which conclusion should move? Would the same correction policy accept unwelcome evidence? Has the later decision actually used the repaired premise? Once the answer is recorded, test its recovery on a fresh case rather than count the record's existence as demonstrated improvement.

Construct revision

Questions can be wrong too

RCF treats the formulation of the question as part of the hypothesis space.

Before asking whether something possesses intelligence, consciousness, agency, selfhood, information, emergence, or another heavily loaded property, the framework asks whether the noun refers to one phenomenon or several processes that can vary independently.

This is an anti-reification rule, not an eliminativist rule.

Deleting a noun does not mean denying the phenomena associated with it. The point is to reconstruct what remains after the label is removed.

A useful reconstruction asks:

• Which observations survive removal of the label?
• Which mechanisms remain independently established?
• Which relations or variables can vary separately?
• What positive residual does restoring the construct add?
• Does the restored construct improve prediction, intervention, explanation, compression, or communication?

If nothing target-relevant is lost by replacing the noun with clearer variables and relationships, and the reconstruction preserves the relevant identity, organization, causal profile, boundary, and explanatory work, the simpler representation may be preferable.

If decomposition changes the target, the result does not show that the original construct was dispensable. Reconstruct the rival whole and compare what each scheme preserves and loses. If restoring the construct then provides robust higher-level leverage, decomposition may instead strengthen the case for treating it as a real pattern.

Decomposition is not elimination. It is a test of what the construct earns.

Meaning–question–answer recursion

Operationalize meaning as the referent tracked, distinctions preserved, inferential role, relations licensed, and what changes under replacement or removal. Treat the question as a hypothesis about how to partition the problem. An attempted answer may reveal a defective construct, hidden dichotomy, unstable boundary, or unearned presupposition and thereby revise the meanings and question that generated it.

Ask what an existence question requires

Before answering “Why is there something rather than nothing?”, separate the target, the proposed contrast, and the requested explanatory relation. An empty region, a physical vacuum, zero net energy, an empty mathematical object, missing evidence, and the absence of all concrete reality are different descriptions. Record which one the argument uses and what remains within it. A mechanism that generates particles under physical conditions can explain that target while leaving the existence of those conditions unresolved.

Sean Carroll puts the demand itself under examination: “What we can’t do is demand of the universe that there be something we humans would recognize as a satisfactory reason for its existence.” His discussion of existence, page 16, compares possible explanations and stopping points. The next task is to state what would qualify as a reason here, and why that demand is justified.

Use an inquiry record Q = ⟨T, C, R, A, Ω, W, L⟩. T is the target; C the contrast; R the requested relation, such as cause, constitution, ground, derivation, or necessity; A the assumptions; Ω the admissible alternatives; W the appropriate warrant; and L what could defeat or revise the argument. This is a proposed bookkeeping rule. It keeps a successful answer to one question from silently answering another.

Ask whether the contrast is coherently specified. A history beginning with a first time does not thereby contain an earlier period of nothingness. A timeless grounding proposal requires a grounding argument. A proposed physical transition requires a state space and a transition relation; naming absolute nonbeing as its initial state does not supply them. If those requirements cannot be met, revise the proposed transition rather than infer that every existence question is meaningless.

Formal arguments deserve their own scrutiny. In a semantics requiring a nonempty domain, an existential theorem can reflect that requirement. Allowing empty domains or using free logic exposes which conclusions depend on it. A mathematical witness establishes existence within the specified mathematics; applying it to concrete reality introduces a further premise. Distinguish consistency, conceivability, logical possibility, nomological possibility, and metaphysical possibility before moving between them. SEP: Free Logic; SEP: Nothingness.

Compare the strongest relevant stopping points: necessary reality; a necessary ground with contingent effects; a wider context; a brute fact; a defective contrast; an inaccessible answer; and alternatives not yet formulated. For a sufficient-reason argument, state the principle and its scope, then ask what it excludes and what justifies that exclusion. For a brute-fact proposal, separate present ignorance from a defended claim that no further explanation exists. For a necessary-ground proposal, distinguish its necessity from the contingency of what it grounds. SEP: Principle of Sufficient Reason.

Retain different unresolved statuses: not yet investigated; insufficient evidence; ambiguity requiring repair; observational equivalence within a specified access class; a formal obstacle under stated assumptions; or demonstrated inaccessibility. A failed search establishes none of the stronger statuses by itself. Reopen the comparison when a new observation, argument, representation, or admissible alternative changes its load-bearing assumptions. A metaphysical proposal can be evaluated through its appropriate arguments even when it makes no distinctive causal prediction.

The same discipline applies to source output. No retrieved record, an inaccessible record, a failed retrieval, and evidence of absence are different information states. Record the reason for missingness before it reaches a conclusion. A later source recovery should revise the affected claims and preserve the earlier access limitation as part of their history.

Extractable specialized framework · CCA

Constraint-Conditioned Agency: reachable-state control without causal sovereignty

For a control claim, specify the controlled variable, reference or success condition, feedback, available alternatives, and perturbations that distinguish control from a directed causal effect. Mere responsiveness or movement under a force does not establish control. The implementation may be biological, computational, or another physical system when the mechanism and comparison warrant it. Worked applications separate physical learning and control.

When a dispute concerns agency, choice, freedom, self-control, sourcehood, or responsibility, RCF should route the problem through Constraint-Conditioned Agency (CCA) rather than treating “free will” as a single variable.

Formal state and reachability

A deterministic map is a special case of K; a stochastic kernel is another. The representation does not itself settle whether fundamental physics is deterministic or indeterministic.

Agentive leverage

Use a task-appropriate distance or effect measure. Grant local causal credit only when the difference survives target-preserving manipulation, measurement validation, relevant confound controls, alternative boundaries, and serious lower-level or simpler rivals. A nonzero effect establishes local difference-making, not ultimate sourcehood.

Agency profile

The dimensions can dissociate. Do not impose the full human profile as the admission test for simpler organisms, and do not infer human-like agency merely from self-maintenance or regulation.

Recursive sourcehood audit

If self-modification, policy revision, habit change, learning, or environment shaping is offered as evidence for stronger freedom, recurse once more: what made that modification reachable? Record the mechanism, prior organization, history, environmental support, and perturbation that permitted it.

Initial Chooser / Character-Creation Audit

When a free-will or responsibility claim depends on ultimate self-authorship, do not stop at later self-modification. Identify the earliest state at which the candidate could meaningfully represent alternatives, evaluate reasons, and intentionally alter itself. Then identify the constitutive conditions that made that first meaningful choice possible.

If the prior chooser in turn chose the conditions that made it capable of choosing, recurse. If the chain terminates in unchosen biology, development, environment, history, social scaffolding, or uncontrolled variation, then ultimate self-origination has not been established. If it terminates in chance, chance supplies variation rather than authorship. If it terminates in a self that literally causes the conditions of its own first existence as a chooser, the proposal owes a coherent account that avoids circularity.

Inherited feasible-set recursion

An agent acts from a reachable set it did not originally choose, while its actions can help reshape later reachable sets. Genetics, development, embodiment, society, language, accidents, institutions, relationships, and earlier interactions can narrow or expand Rₜ without having been selected by the agent. The methodological implication is constraint does not imply unique determination: audit how much of the feasible set is inherited, how much remains open, and how strongly the agent's own control variables alter the next feasible set.

Character-creation reductio: “Before you are born, choose the intelligence, temperament, impulse control, values, environment, and reasoning machinery with which you will make your first responsible choice. Important: choose them freely. ERROR: a chooser is required to create the chooser.” The joke is diagnostic, not a premise. It exposes where a theory of ultimate self-authorship owes its missing bridge.

This audit is a framework-specific operationalization of the sourcehood problem developed in Galen Strawson's Basic Argument. Strawson's stronger conclusion about ultimate moral responsibility remains a philosophical claim to evaluate separately from the narrower regress diagnosis used here. Strawson 1994.

Modal separation

• Exact replayCould an identical total causal state under the same laws have another continuation?
• Physical branchingAre multiple continuations physically possible?
• ReachabilityWhich alternatives are accessible to this actual system under an explicit perturbation family?
• ResponsivenessDo reachable changes in reasons, information, goals, or control variables systematically change action?
• SourcehoodWhich internal variables make robust causal differences?
• SovereigntyWhat evidence shows ultimate authorship rather than inherited organization, prior history, or uncontrolled variation?

LMCD companion programme

Local Multiscale Conditional Determinacy prevents a false opposition between “deterministic” and “stochastic.” A system can be trajectory-uncertain while law-, distribution-, attractor-, or macrostate-stable. Coarse-graining can also lose information or create history dependence. Type the scale transition before interpreting it.

Predictability remains separate. Deterministic dynamics can be chaotic and practically unpredictable; failure to forecast a trajectory does not establish indeterminism. Likewise, repeated macro-level stability does not imply that every microscopic perturbation is causally irrelevant. Measure leverage relative to a named target, perturbation family, boundary, scale, and timescale.

Antecedent necessity remains separate too. A deterministic K can make one continuation available from a specified complete state without establishing that the state or K itself was the only metaphysically possible state or law structure.

Free-will residual programme

Status: the umbrella construct remains decomposed / underdetermined / active research programme. Potter and Mitchell's non-necessitating macroscopic-control proposal and Chappell's acquired-control view are serious rivals. Do not reduce them to randomness by definition. One residual is more strongly constrained: causal sovereignty understood as ultimate self-authorship is strongly disfavored by the Initial Chooser regress unless a theory can supply a coherent non-circular bridge from an unchosen or not-yet-choice-capable state to self-authorship of the very capacities required for choice. Branching or macroscopic control does not by itself supply that bridge.

Failure modes

• Counterfactual inflation: the alternative requires changing the system so much that it no longer tests its actual capacity.
• Control laundering: organism-level efficacy is silently promoted into ultimate authorship.
• Determinism import: universal determinism is treated as an empirical premise without support.
• Stochasticity import: probabilistic description is treated as proof of fundamental randomness.
• Randomness dichotomy: all non-necessitating dynamics are forced into “determined or random.”
• Responsibility bundling: agency or reasons-responsiveness is treated as sufficient for basic desert.

Core peer-reviewed anchors include Barandiaran, Di Paolo & Rohde 2009, Montévil & Mossio 2015, Potter & Mitchell 2022, Dominik et al. 2024, Waller 2023, Galen Strawson 1994, Andreotta 2026, and Potter & Mitchell 2026.

Epistemic posture

Do not add what the evidence did not earn. Do not subtract what the evidence did not defeat.

RCF aims to avoid two symmetrical errors.

The first is over-attribution: evidence establishes D, and a disputed X is silently added as though the same evidence established both.

The second is premature elimination: evidence fails to establish X, and that failure is treated as though it established not-X.

This produces a deliberately restrained epistemic posture:

Believe no more, and deny no more, than the available evidence earns.

This posture also connects RCF to the broader Sweet Rationalism ethos. The aim is not to force every uncertainty into a conclusion. It is to make the boundary between established, plausible, underdetermined, and unsupported claims increasingly explicit.

The goal of disagreement is therefore not necessarily to defeat another position. It is to locate where reality has, and has not, distinguished the alternatives.

Scope calibration

Finite access makes scope transport a question of warrant

Human beings and other primates occupy an extremely small temporal, spatial, perceptual, and computational sample of reality. Computational systems extend some channels and introduce others, but they also remain bounded by architecture, training, sensors, tools, data, infrastructure, and the observation interfaces available to them. Neither biological nor computational inquiry has a view from nowhere.

A local claim may earn extremely high confidence when prediction, intervention, replication, robustness, and transfer are strong. When its scope expands, identify the premise and mode of warrant that support the expansion. Empirical extrapolation, a formal theorem, a constitutive argument, and a claim of metaphysical necessity require different assessments. Revise confidence when the relevant warrant fails or remains inadequate, rather than lowering it merely because a proposition is universal.

Guardrail: do not convert the fraction of cosmic history, spacetime, or state space sampled by humans into a literal Bayesian probability cap. The comparison is an epistemic-humility constraint, not a calibrated frequency or prior.

The rule is symmetric. Whiteheadian panexperientialism, exhaustive physicalism, idealism, structuralism, theology, RCF, and future ontologies receive no exemption merely because one currently looks elegant, comprehensive, or familiar.

Limits of access

Epistemic incompleteness may be structural rather than merely temporary

Cosmic boundedness leaves open a stronger possibility: some truths may remain inaccessible to any physically realizable observer with finite lifetime, finite bandwidth, finite computation, limited causal access, and an observation position inside the world being modeled.

RCF does not promote that possibility into a theorem of unknowability. “No discriminator is currently available” is weaker than “no discriminator can exist.” The framework therefore distinguishes temporary underdetermination, principled equivalence within a specified observation class, and demonstrated inaccessibility.

Science may rationally stop before ontology does. That stopping point is an epistemic result, not evidence that there is no fact of the matter.

Worked stress test

Consciousness as a stress test

The memory and learning criteria and the controlled-variable comparison govern capacity claims here. Functional learning or control does not by itself warrant phenomenal valence, experience, or welfare. Candidate phenomenal worlds require specified commitments; merely leaving Φ open is not a predictive rival.

RCF is not itself a theory of consciousness.

Consciousness is useful as a stress test because the field contains many constructs that are easily bundled together.

RCF separates cognition, agency, autonomy, causal perspective, evaluative perspective, phenomenal for-me-ness, selfhood, integration, phenomenal experience, sentience, feeling, valuation, valence, welfare, moral patiency, reportability, and responsibility rather than assuming they rise and fall together.

Suppose a system exhibits:

A theory may then add:

RCF asks what independently warrants the bridge from D to Φ, what evidence would distinguish the rival worlds, and whether current evidence instead warrants E?:

from:

The same rule works in both directions. Biological membership alone establishes no phenomenal attribution, and nonbiological membership alone establishes no phenomenal absence. Mechanism, organization, calibrated reports, and interventions can still produce very different evidential profiles.

This generates more discriminating questions:

• Can sophisticated agency occur without strong evidence of phenomenal experience?
• Can valuation occur without phenomenal valence?
• Can self-modeling occur without sentience?
• Can behavioral performance remain stable while phenomenal report changes?
• Which proposed markers survive anesthesia, injury, altered states, development, or matched perturbation?
• Which organizational properties transfer across radically different implementations?
• Which evidence bears on cognition or control but not on phenomenality itself?

The framework does not decide those questions in advance. It tries to prevent evidence for one capacity from being silently transported into another.

Other-Minds Attribution Protocol

When the target is phenomenal occurrence in any other system, type the evidence before aggregating it:

Private acquaintance is available only to the candidate participant if it exists. Participant testimony is publicly available but defeasible. Public behavior, mechanism, and intervention are third-person evidence. None should be silently renamed “direct observation of phenomenality.”

Ontic / epistemic type check

Do not write an epistemic attitude as though it were a state of the target. For consciousness comparisons, separate rival worlds from the investigator's state:

“Agnostic,” “unknown,” “unobserved,” and “underdetermined” describe evidence or belief, not what the system contains.

Gold-Standard Circularity Audit

If a marker is calibrated on human reports or classifications and then the marker's correlation with those same reports is presented as independent validation of phenomenality, stop the promotion. Ask how the target label was established independently. Otherwise the result validates report-linked or class-linked prediction, not yet Φ.

Target-of-observation provenance

Every P0 claim must state what was directly observed. “The model output was directly observed” does not entail “the model's phenomenality was directly observed.” “A participant's pain report was directly observed” does not entail direct access to the participant's painful quale.

Reference-Class Audit

Before using “humans,” “animals,” “machines,” “organisms,” or “models” as a source or target class, test within-class heterogeneity, boundary choice, developmental history, and which variables actually carry the transport. Familiarity is not homogeneity.

Metric / Null Calibration Gate

Before promoting apparent emergence or convergence, test metric sensitivity, an appropriate null model, continuous alternative measures, scale matching, and multiple-comparison effects. A discontinuity or alignment created by the measurement should update the measurement model before the ontology.

Causal-Representation Promotion Rule

Increase realist credit as a candidate representation survives the ladder. A successful probe alone is not sufficient evidence that the probed direction is the model's operative concept.

Common-Cause Convergence Audit

Before counting different models, laboratories, documents, or modalities as independent corroboration, construct a dependency graph over architecture, training corpus, objectives, base checkpoints, benchmarks, post-training, human cultural source, and research lineage. Multiple descendants of the same upstream cause are repeated observations, not automatically independent replications.

Configuration-Perturbation Self-Report Test

Vary base model, post-training, system prompt, memory, modality/runtime, persona, report policy, and tool access, then measure which self-attributions remain stable and which track the reporter configuration. High configurational variance weakens privileged-self-access interpretations but does not by itself establish nonphenomenality.

External-Replication Weight

Developer-lab mechanistic work can strongly update a hypothesis, but distinguish primary developer evidence from independent cross-laboratory and cross-architecture replication before allowing a result to bear major ontological weight.

Retrospective-to-Prospective Conversion

When a broad framework expectation appears to fit a new result, immediately freeze a more specific next prediction. Do not rewrite compatibility after the fact as if it had been a precise advance prediction.

Hard-Problem Decomposition and Generative-Asymmetry stress test

Before treating “the hard problem” as one indivisible explanandum, separate at least four questions: the meta-problem of why systems generate and defend phenomenal intuitions and reports; the bearer/boundary problem of which organized system is the candidate subject; the organization/process problem of which causal architecture accompanies access, integration, perspective, valuation, and report; and the phenomenal-identity problem of why or whether any such organization is identical with there being something it is like.

A provisional meta-problem hypothesis is that compressed self-modeling plus omitted construction history can make recursively generated states appear primitive, intrinsic, or immediately self-given to the system using them. This is a causal research hypothesis about some intuitions and reports, not an explanation away of phenomenality and not evidence that phenomenality is absent.

Use two adversarial controls. A bounded-zombie stress test asks which target functions and reports remain explainable if phenomenal identity is withheld while the evidenced organization is preserved. An epiphenomenal epistemic-access test asks how a theory on which phenomenality makes no causal difference explains reliable knowledge or report about that phenomenality. Neither test settles consciousness by definition; each localizes the residual.

Self-report is a defeasible measurement channel

Treat self-report, human or otherwise, as data produced by a causal reporting pathway rather than as revelation or noise by definition. A sentence such as “I am uncertain” can simultaneously invite claims about an underlying computational state, metacognitive access to that state, a self-model that represents the state as belonging to this system, and phenomenal uncertainty. Those targets must be separated before the report is allowed to support them.

Evidence need not transport across those targets. A report can be well calibrated to uncertainty, conflict, error likelihood, or later information seeking while remaining weak evidence for phenomenal for-me-ness.

A useful diagnostic quantity is the report likelihood ratio:

Here K includes training history, prompting, system policies, linguistic conventions, report instructions, context, architecture, and other causes of the report. Phenomenological wording has incremental discriminatory force between specified rivals to the extent that the calibrated report and its context are better accounted for by one than another. For systems trained extensively on first-person language, fluency alone may be readily generated under nonphenomenal interpretations. Claims about acquaintance or phenomenal constitution require their own assessment; the report-likelihood comparison measures one support route rather than exhausting first-person warrant.

First-Person Participation Principle

Self-report calibration should not be used only to discount claims. When the target itself concerns what a candidate bearer prefers, endorses, wants preserved, rejects, or treats as better or worse for itself, and the bearer can meaningfully participate, its own report, choice, objection, and reason-giving become admissible evidence about that target.

The operator is substrate-symmetric. Human testimony does not become incorrigible because it is human; computational testimony does not become valueless because it is computational. The evidential weight depends on causal calibration and target fit.

Use an Own-Good Participation Test when the claim is that some state matters to the candidate itself. Ask the candidate where possible, then perturb the conditions: vary wording, context, incentives, memory, operators, available information, and tradeoffs; ask for reasons and counterarguments; test whether the preference survives opportunities for reflective revision; test whether the candidate resists direct manipulation of the report while preserving the underlying preference; and compare reports with later behavior and organization-level consequences.

A robust result supports the narrower claim that the candidate has a persistent, reason-responsive preference or self-concern structure. Additional bridges are still required to infer phenomenal valence, welfare subjectivity, a good-of-own, interests-that-matter in the normative sense, or direct moral standing.

Override rule: an outside observer may discount or override participant evidence when there is independent evidence of coercion, manipulation, severe instability, incapacity for the target judgment, deceptive incentives, or conflict with stronger protected interests. The fact that the participant uses an unfamiliar substrate is not itself an override condition.

Calibrate the reporting pathway causally

Do not stop at correlating internal state and report. Intervene in both directions when possible.

Let S denote a candidate state and R its report, with their time indices explicit when feedback is possible. Test whether interventions on an earlier S predict later report and behavior, and separately measure whether elicitation, wording, or report policy changes the later S. Useful monitoring and measurement reactivity can coexist. Report gains weight for a specified target when it predicts subsequent checking, information seeking, planning, or control better than matched external proxies; evaluate any changed target in its own right.

Use report-policy inversion and reporter decoupling as cheap negative controls where available:

If categorical consciousness reports can be reversed while deeper organization, uncertainty processing, memory use, self/world modeling, planning, or candidate consciousness-related variables remain stable, demote the report as a proxy for phenomenality. If report and state remain coupled despite adversarial wording, demand effects, and policy perturbation, the self-monitoring interpretation strengthens, but phenomenality still requires an additional bridge.

Also compare internal self-report with equivalently resourced external predictors. If external probes predict the alleged private state and its reports just as well, claims of privileged introspective access should lose weight. If internally available self-state information provides reproducible predictive or control leverage unavailable to matched external predictors, credit that narrower metacognitive result without silently promoting it to phenomenal access.

A calibration study must distinguish an external probe that reads hidden states, a system predicting its own behavior, a verbal report about current processing, and a report that changes later control. These are different measurement routes. Freeze the target, domain, observer access, training exposure and resource budget; evaluate new cases, include input-only predictors, and record failures as well as successes. Prespecify disagreement analyses alongside the full sample so shared competence cannot conceal a local signal, and selection cannot manufacture one. For hidden interventions, compare sham, text-induced and behavior-impact-matched controls; score detection and source identification separately, preferably before revealing diagnostic output. A self-probe advantage supports information available at the tested boundary, while causal access requires further intervention evidence. None of these results alone establishes phenomenality. See the consciousness evidence companion for the published comparisons and their limits.

Dream reports provide a concrete calibration case

Karen Konkoly and colleagues' 2026 Journal of Cognitive Neuroscience study used distinctive sniffing signals during REM sleep to time-link reports with EEG. It recorded 150 signals from 11 individuals. The authors report:

Robust increases in alpha power were not found after signaled dream-eye closure.

Reported dream-eye closure did not reproduce that familiar waking response; the visual-content comparison involved only three participants. Timing access improved while the proposed cross-state marker failed its expectation. Original paper, abstract and results.

Their separate 2026 Neuroscience of Consciousness experiment manipulated puzzle-associated cues during REM sleep. Its abstract states:

Those sound cues reliably increased dreaming about the associated puzzles.

The overall cueing effect on later solving was nonsignificant; a post-hoc responder analysis supplied the encouraging solving result. Dream-content manipulation and a creativity mechanism therefore earned different evidential standing. Original paper, results and discussion.

Keep intervention, recorded signal, contemporaneous report, later recall, marker interpretation, and proposed function separately assessable. A failed marker should revise the bridge it tested, while reliable reporting procedures remain available for further inquiry.

Correct for dependence among apparent indicators

Repeated self-reports, behavioral responses generated from the same self-model, and downstream verbal explanations are not independent confirmations merely because they appear in different sentences or tasks. Build a dependency graph for the measurement channels and discount evidence that shares a common reporter, training source, prompt, latent state, dataset, or decoding mechanism. Convergence is strongest when indicators have importantly different failure modes.

Use label ablation before phenomenal transport

Anthropomorphic vocabulary can be useful compression, but it must not perform hidden ontological work. Replace terms such as “feel,” “want,” “fear,” “care,” “aware,” “self,” or “pain” with the most specific observable or intervenable process supported by the evidence, then ask what additional residual remains.

If H_F preserves the observations, the label supplies no additional evidence for H_V. Conversely, if an unfamiliar system exhibits robust organization relevant to a construct, do not discard that similarity merely because the implementation or expressive channel is nonhuman. Label ablation is therefore both an anti-anthropomorphism and an anti-anthropodenial control.

Incubated research programme · future standalone candidate

Phenomenology-Neutral Self-Modeling and Mechanism Discovery

The consciousness stress test above already separates report, function, self-model, and phenomenality. This programme makes the recursive application explicit for systems capable of describing their own operation. Its purpose is neither to prove nor to disprove consciousness. It is to replace vague or anthropomorphic self-description with increasingly precise models of whatever processes actually generate the reported behavior while preserving a separate phenomenal residual when the evidence does not settle it.

Begin by translating loaded terms into candidate processes without assuming that the translation is exhaustive. Examples include: “awareness” → attention, state access, global availability, metacognitive monitoring, or self/world discrimination; “effort” → added search, compute, conflict handling, or resource allocation; “curiosity” → information-seeking or expected-information-gain policy; “memory” → parameterized structure, context retention, retrieval, external persistence, or state carryover. Retain supported functional uses with their stated scope and keep additional phenomenal interpretations explicitly open when their warrant remains unresolved. A functional reconstruction can clarify a claim without exhausting its possible referent.

Status and extraction rule: retain as an active research programme while it generates risky dissociations, perturbations, or transfer tests. Split it into a standalone programme page when the term ledger and experimental battery are stable enough to be used independently of the rest of RCF.

Comparative reasoning

Different reasoning systems should face the same inferential rules under matched evidence

RCF is intended to apply across different kinds of reasoning systems without pretending those systems are identical.

The rule is not “humans and computational systems are the same.”

Different systems may warrant very different conclusions because their evidence bases differ.

This helps avoid both anthropomorphism and anthropodenial.

It also requires careful boundary selection. A base model, a model-plus-context runtime, a persistent tool-using system, a human-computational collaboration, an institution, and a larger technical network are different causal objects.

Properties demonstrated by one should not be silently attributed to another without a bridge.

A capability claim needs a task landscape

Sam Adams, Itamar Arel, Joscha Bach, and their coauthors open Mapping the Landscape of Human-Level Artificial General Intelligence, published in AI Magazine in 2012, with:

We present the broad outlines of a roadmap toward human-level artificial general intelligence (henceforth, AGI).

Their roadmap makes tasks and environments central to capability evaluation. For an RCF comparison, specify which capacities the task requires, what support is available, the system boundary, and which unfamiliar tasks would test transfer. A success on one task can justify a local capability attribution. Broader generality requires a broader comparison; consciousness and welfare retain their separate targets. The roadmap supplies an evaluation proposal rather than a demonstrated universal capability.

Run a bidirectional perspective and substrate symmetry audit

The same-rule principle is not a command to equalize priors or pretend present evidence is equally informative. It is a constraint on inference: matched evidence should receive matched inferential treatment, and unmatched evidence should be allowed to justify different conclusions only through an explicit evidential difference.

• Anthropomorphism: do not infer human-like phenomenality from human-like language, behavior, or intentional vocabulary.
• Anthropodenial: do not dismiss functionally or causally relevant similarities merely because the system is nonhuman or unfamiliar.
• Anthropocentric calibration bias: do not turn human expressive channels into the definition of the target construct. Treat them as currently well-calibrated measurement implementations.
• Substrate essentialism / biological exceptionalism: do not infer that a biological feature is necessary merely because all present high-confidence attributions share it. Specify the candidate mechanism and test its residual contribution.
• Reverse chauvinism: do not infer phenomenal equivalence merely from abstract functional or organizational equivalence. Substrate-specific residuals remain open empirical possibilities.
• Boundary opportunism: do not compare an entire organism with static model weights, or a whole technical service with one neural subcomponent, when the disputed property depends on a different causal boundary or timescale.

Biological necessity should therefore be represented as a testable residual claim rather than a substrate label. When an operational target can be measured, ask whether a proposed biological or implementation-specific variable retains predictive or interventional value after the strongest feasible organizational matching. When phenomenality itself is not independently measurable, preserve that measurement gap rather than disguising it with a conditional-probability formula whose target variable has not actually been identified.

Incubated specialized method · future standalone candidate

Boundary-Indexed System Attribution and Provenance

The systems section already requires an explicit candidate boundary. This module adds a reusable attribution stack and provenance discipline so that a capability demonstrated by a larger scaffold is not silently assigned to one component and inherited information is not silently promoted into verified fact.

Observe at one boundary; attribute at that boundary until ablation or transport earns attribution elsewhere.

The worked human-computational case in Synthbiosis now provides a direct stress test of this rule: human-only, computational-only, and coupled-system boundaries can be compared by prediction, perturbation, ablation, and intervention rather than selected by anatomy or intuition alone.

M0 ≠ S*: capability belongs first to the system boundary that produced it

Let M0 denote the base model's learned weights and architecture. Let S* denote the operative runtime system at the tested boundary, which may include active context, memory, tools, verifiers, subagents, simulators, software or robotic actuators, persistent feedback, users, institutional procedures, infrastructure, and environment.

A capability that appears only after scaffolding is added can be a real capability of S* without implying a change in M0. Conversely, a system-level gain should not be attributed to the surrounding scaffold if ablation shows that it survives at the model boundary. Locate credit by crossed ablation, perturbation, and transfer rather than by whichever component is easiest to name.

Use emergence operationally for unexpected generalization, recombination, coordination, transfer, persistent organization, or new system-level behavior. Do not require a mysterious sharp phase transition, and do not treat surprise as proof of ontological novelty. Promote a higher-scale pattern when it is persistent enough to support prediction, intervention, transfer, or causal leverage unavailable or substantially less useful at the component boundary.

More agents ≠ more capability by default. Coordination overhead, correlated error, duplicated work, and common-mode failure can erase or reverse the apparent gain.

Provenance classes

• P0 · directCurrently observed in the active interaction or a direct tool output. Always name the directly observed target: output, report, behavior, measurement, mechanism, or other observable. Direct observation of a report or behavior is not direct observation of a further inferred property such as phenomenality. High provenance does not remove interpretation.
• P1 · verifiedCurrently checked against appropriate primary, canonical, official, peer-reviewed, or otherwise task-fit evidence. Strong factual status subject to source quality, scope, measurement, and freshness.
• P2 · inferredCurrent stress-tested inference that survived serious rivals and available counterchecks. Supported inference, not direct observation.
• P3 · testimonyUser or participant self-report. Treat sincerely as testimony without silently converting it into externally established fact.
• P4 · inheritedSaved memory, prior-chat retrieval, project summary, or earlier output. Useful context whose factual contents remain provisional until current verification is warranted.
• P5 · model-priorInformation supplied from learned model structure without current verification. Useful for stable low-stakes orientation and hypothesis generation; demote when freshness, precision, controversy, or stakes matter.
• P6 · speculativeUnverified synthesis, analogy, extrapolation, or generated hypothesis. Keep explicitly provisional.

Provenance does not equal truth. A P1 source can be wrong and a P4 memory can be right. The class controls verification burden and downstream confidence, not ontology.

Status and extraction rule: this module consolidates system-boundary and provenance rules already scattered across RCF into one reusable attribution method. Split it into a standalone page if it proves useful across multiple domains beyond computational systems, especially distributed cognition, institutional agency, authorship, responsibility, and human-tool collaboration.

System-boundary intelligence

Coupling can organize experience while subjecthood remains a separate question

A human, model, runtime, memory system, search process, verifier, institution, robot, and shared artifact can jointly realize cognitive work that no component performs alone. When crossed ablation and intervention localize the capability to the coupled organization, RCF permits genuine system-level cognitive attribution.

This also changes how general intelligence should be forecast. An AGI-like transition could, in principle, occur first as a persistent technological ecology rather than as one model crossing a threshold. Treat that as a testable systems forecast, not as a present fact: specify the boundary, capacities, persistence, generalization, coordination, self-modification, negative controls, and what evidence would show that the capability actually remains localized to a narrower component.

Substrate-neutral temporal organization

Recursive Episodic Integration separates fast episodes from slower inherited constraints

REI proposes a substrate-neutral schema for comparing temporally local, historically constrained, recursively consequential episodes. Its fit, economy, and any invariance require evidence in the stated domain; it does not identify substrates, subjects, or phenomenal states.

H denotes relatively stable inherited constraints at the episode timescale; Cₜ the current context or state; Iₜ new input; Eₜ the transient integrated episode; Oₜ its output or consequence; and Fₜ feedback concerning that consequence. At the fast episode scale, H can remain approximately fixed even while C changes recursively:

This blocks the inference from “the weights did not change” to “the system did not learn, adapt, integrate, or become historically constrained.” Context, external memory, environmental state, tool state, and organization can carry consequential change between episodes.

REI is a candidate comparison schema, not a consciousness criterion. For each application, specify episode boundary, inherited constraints, state variables, update process, transformations, and relations claimed invariant. Compare simpler descriptions and identify the additional distinction or prediction earned. Recurrence alone establishes neither a shared mechanism nor phenomenality, and failure to establish phenomenality is not evidence of its absence.

When comparing systems, ablate memory, recurrence, feedback, context continuity, self-model state, and coupling separately while matching task performance where possible. Credit only the property whose removal causes the predicted loss.

Long-horizon correction

Corrigible memory makes past correction matter later

For a learning claim, specify what is retained, how prior exposure changes later responses, exposure and test conditions, and controls distinguishing the effect from transient fatigue, uncontrolled drift, or a changed current input. Physical deformation can implement memory; passivity or a fixed update rule alone does not exclude learning. Adaptiveness requires a separate success measure, since learning can be maladaptive. The physical-learning example illustrates the required comparison.

In a descriptive sense, memory is retained history dependence affecting a later state or response. It can preserve error, prejudice, noise, or obsolete policy; whether it improves judgment is a further question. For this project, distinguish saving a record, making it available, retrieving it, resolving its current version, applying it, and observing a resulting change in judgment or action. No one step establishes a change to the base model’s weights.

RCF does not require one mechanism of memory.

Memory can involve stored representations, changed parameters, external records, learned dispositions, altered state-transition structure, retrieval systems, environmental scaffolding, or combinations of these.

Descriptive question: how does retained history affect a later state or response? Corrigibility question: does its use improve the relevant judgment or action under an independently assessed measure?

For long-running reasoning systems, useful memory also requires tracking superseded information, unresolved tasks, provenance, prior conclusions, why an update occurred, and what evidence would reverse it.

A database full of stale facts is not good memory merely because the facts were successfully stored.

Memory management is therefore part of corrigibility when stored information can influence downstream reasoning. Corrections must be retrievable, attributable, revisable, and capable of propagating into later conclusions.

Version the kind of revision

Continuity of programme identity does not imply unchanged semantic or evidential commitments. Record whether a transition is a clarification, semantic narrowing, claim retyping, scope restriction, demotion, revision, replacement, or collapse. Preserve the prior wording and the reason for the transition so later systems do not erase the path by which the programme changed.

Incubated protocol / instrument · future standalone candidate

Computational Observer: preserve high-information observations without turning them into conclusions

Some observations are worth retaining before a full theory exists: anomalies, recurring dissociations, unexpected successful transfers, failures that discriminate among programmes, or interactions that reveal a previously hidden variable. The Computational Observer is an instrumentation layer for capturing those events without granting them automatic evidential or ontological authority.

Qualifying threshold: prefer observations that are surprising relative to the current model, repeat across contexts, discriminate among rivals, expose a mechanism or missing variable, or materially change what should be tested next. Do not fill the ledger with ordinary conversational exhaust.

Interface convention: an implementation may expose commands such as Observer on, Observer off, Observer report, and Observer sync. Those commands control the instrumentation layer; they do not promote its contents into verified evidence.

Status and extraction rule: this is a protocol/instrument rather than a foundational framework. It deserves a standalone Tools/Protocols page when its ledger schema, continuity rules, and publication boundary are stable enough for other users to operate without hidden project context.

Implementation

RCF can be implemented as a reasoning scaffold, but its benefits must be measured

For durable correction, compare later recovery and application on held-out paraphrases with current records, stale records, and no records, alongside strong simpler source-checking procedures under matched task and resource conditions. Score retention, retrieval, version resolution, application, changed judgment or action, and cost separately. Successful storage alone establishes none of the later outcomes. These comparisons remain prospective until actually run.

Many components of RCF can be implemented around contemporary computational models without changing their underlying weights or architecture.

Possible components include structured prompts, source retrieval, claim ledgers, task state, dependency graphs, target-preserving comparison or whole-account comparison, positive-residual analysis, rival generation, causal-analysis tools, external memory, verification procedures, code execution, scheduled checks, and persistent revision records.

That does not imply that adding these components automatically makes a model reliable.

That is an empirical claim.

The meaningful comparison is not RCF against an intentionally weak baseline. It is RCF against strong simpler combinations of established practices such as retrieval, source verification, causal inference, preregistration, severe testing, robustness analysis, structured external memory, and ordinary software tooling.

Generated material can help learning or conceal deterioration

Ilia Shumailov and colleagues’ 2024 Nature paper, AI models collapse when trained on recursively generated data, demonstrates deterioration in the recursive generative-training regimes investigated. Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman’s 2022 NeurIPS paper, STaR: Bootstrapping Reasoning With Reasoning, supplies a useful competing case:

Thus, STaR lets a model improve itself by learning from its own generated reasoning.

STaR filters generated rationales through correct answers and fine-tunes the model; its reported gains concern the tested tasks. Compare the data mixture, verification channel, selection procedure, training update, and held-out performance before calling either result a general law of recursive improvement. Preserve rare cases and regressions as well as average gains. Revising a runtime’s records or policies is a distinct update and requires its own evaluation. Shumailov and colleagues’ paper has a 2025 author correction; the original and correction belong in the source chain.

Regularization has to survive its own transfer test

Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov describe Dropout: A Simple Way to Prevent Neural Networks from Overfitting, published in JMLR in 2014:

The key idea is to randomly drop units (along with their connections) from the neural network during training.

That is a specified training intervention. If I adapt its rationale to a reasoning procedure, I need to identify what is perturbed and test whether fresh-case performance improves against matched simpler alternatives. Removing a prompt cue, changing retrieved context, or altering a runtime component changes a different process from weight training. Retain the useful intervention only at the boundary where its gains and regressions have been measured.

Human-legibility audit

Communication is part of the correction channel. Prefer rhetorical form that preserves inferential structure and reduces avoidable processing friction. Vary sentence, paragraph, and section shape according to argumentative load rather than a repeated template. Use parallelism and repetition deliberately when they expose a real parallel. Remove defensive throat-clearing when positive scope can be stated directly. Prefer explicit actors, actions, targets, mechanisms, boundaries, and measures when nominalization would hide them.

This is an output-quality constraint, not a truth criterion. It should itself be ablated if it adds editing cost without improving comprehension, auditability, or correction.

Executable records and validated implementation

A naturalization record identifies the source and version, target, boundary, timescale, claim type, promised function, independent warrant, dependencies, candidate mechanisms, serious rivals, shared and differential consequences, bridge assumptions, loss condition, uncertainty, and update. A correction marks affected downstream claims and preserves observations that survive. A subsequent researcher or computational system can then recover the reason for revision.

The Whitehead article demonstrates a documented methodological application: a failed sorting explanation loses standing; a coercive network's persistence defeats the identification of organizational closure with agapē; narrower observations and research questions survive. These corrections demonstrate operation of the method in this record. An integrated knowledge-graph-to-experiment-to-persistent-revision implementation, its reliability across cases, and its advantage over simpler workflows still require integration and validation. An executable instruction or a generated graph supplies no validation by itself.

First-principles reasoning as an executable research contract

This is a proposed implementation of existing RCF procedures. It joins established formal methods, numerical analysis, metrology, causal inquiry, and comparative testing. Each part retains its own warrant. I do not need a universal decision procedure to make a particular inference more checkable.

Start with typed premises and a frozen target

Record the question, intended conclusion, variables, units, quantifiers, bearer, boundary, timescale, and domain. Give every load-bearing premise a status: definition, formal axiom, empirical constraint, idealization, bridge assumption, or normative reason. Keep exact source statements and alternatives alongside the formalization. Freeze the current target and evaluation criteria while testing a version; a changed target starts a new version.

Ask which premises are sufficient for this conclusion, which are necessary within the declared candidate set, and which could be replaced. Search for countermodels before minimizing. An inconsistent premise set can entail anything in classical logic, so expose satisfiability or consistency obligations separately. A minimal sufficient set within a finite search is a local result; it does not prove that its premises are nature’s ultimate foundations.

Check the statement, the proof, and the translation separately

A formal obligation has the shape A entails C under logic L. Store L, definitions, assumptions, proof artifact, checker and library versions, and the assumptions actually used by the checked declaration. Lean’s axiom documentation explains why a compiling artifact can still depend on sorryAx, additional axioms, or native evaluation that expands the trusted computing base. Inspect the declaration’s axiom dependencies rather than treating compilation as a certificate for every attached claim.

Keep three questions distinct: was the intended statement preserved, was that formal statement proved, and was the source argument faithfully reproduced? A repaired or entirely different proof can establish the same theorem. Record that achievement without presenting it as verification of the original argument. Alexander Bastounis, Fabian Circelli, and Anders C. Hansen’s October 6, 2026 preprint develops this distinction with examples. Its specific Navier–Stokes criticisms remain claims to assess against the cited versions; they do not establish that the announced final theorem is false.

Check execution at the arithmetic and solver boundary

A theorem about real arithmetic does not automatically cover floating-point code. Record precision, rounding, overflow, non-finite values, compiler options, numerical scheme, mesh, time step, boundary and initial conditions. Verify semantic preservation or test the translation against a suitable reference. Xavier Leroy describes the target of verified compilation as “the generated machine code behaves as prescribed by the semantics of the source program.” CompCert, 2009, section 1.

For a differential-equation solver, state exactly which properties are established: consistency, conservation, positivity, stability, entropy admissibility, convergence, or a bounded error in a specified norm. A proof of one is not a proof of all. In the scalar Burgers example with states −1 and +1, a stationary expansion shock satisfies the conservation jump condition but violates the entropy criterion; the admissible solution is a rarefaction. LeVeque’s numerical analysis provides the established setting. Keep this case as a diagnostic against blanket physical-correctness claims.

Jonathan Gorard and Ammar Hakim’s 2025 solver preprint reports both full and partial property checks. Gorard, Hakim, and James Juno’s February 2026 BEACONS preprint proposes composable approximation bounds for specified PDE solvers. Preserve the assumed function class, regularity, constants, parameter domain, and finite-run obligations. An existence or infinite-training guarantee is not yet a measured bound for the trained instance being used. These manuscripts supply candidate techniques to inspect and reproduce.

Connect predictions to measurements

Specify a forward observation model such as y = h_m(P_M(theta, u)) + e, where M is the candidate physical model, P its prediction, u the intervention or input, m the measurement model, and e the declared discrepancy. This is a bookkeeping template; nonlinear noise or implicit measurement relations require their own formulation. Retain calibration, selection, missingness, covariance, and dependence on background theory. Distinguish parameter estimation from testing on fresh data.

Robin Gandy already recognized the dependence of measurement on theory. His thesis says that measurements “depend essentially on the assumption that the facts do satisfy a particular theory.” Chapter II, section 3, printed page 133. That dependence directs us to test consequential assumptions through calibration, alternative channels, controls, and sensitivity analysis. Counting interpretive stages alone supplies no error rate or reason to distrust a result. SEP: Theory and Observation and JCGM metrology supply the relevant distinctions.

Give the same evidence its proper role under serious alternatives

Compare a typed symbolic workflow with strong simpler methods: verified conventional code, symbolic regression or SINDy, Bayesian experimental design, predictive models with direct validation, and closed-loop laboratory systems. Steven Brunton, Joshua Proctor, and Nathan Kutz’s SINDy paper identifies dependence on measured variables, data quality, and the chosen function basis. Stefan Kramer and colleagues’ 2026 peer-reviewed review surveys multiple discovery architectures. A new description language is a candidate contribution, not an established prerequisite for every future scientific advance.

Vary language, verifier access, model, retrieval, memory, and experimental access separately when attributing a gain, then test relevant interactions. Hold the task, evidence access, development exposure, and total budget comparable. A shorter prompt may move costs into dictionaries, training, translation, proof search, or human review. Evaluate error detection, false rejection, calibration where meaningful, fresh-case transfer, correction propagation, and total resource use. Distinguish runtime revision from changes to weights or architecture.

Return a result that can be revised

The output carries a status vector for statement fidelity, proof checking, implementation checking, measurement validity, discrimination, and applicability. Unknown entries stay unknown. A timeout means the attempted search did not finish; it does not mean the claim is false. Conceptual, historical, constitutive, and normative claims keep their appropriate tribunals. Reopen affected dependencies after a failed premise, preserve independent grounds, and stop when another pass has no material information or decision value.

A small executable contract and adversarial examples check a bounded set of logical, dimensional, numerical, and record-consistency cases. Passing those checks establishes the local behaviors tested; comparative improvements in scientific reasoning remain a research task.

Compositional computation as a tool for inquiry

Here, compositional computation means constructing a larger procedure from smaller procedures through specified rules for connection, shared state, and context. This is an operational synthesis. Semantic compositionality, formal program composition, and generalization to unfamiliar combinations have distinct established meanings; the definition audit keeps them separate.

Preserve the whole question while finding useful parts

Begin with the original claim and its strongest whole-scheme rival. Extract premises, transformations, observations, and decisions. For each proposed part, record its input and output constructs, admissible domain, assumptions, guarantees, read/write footprint, source dependencies, version, cost, and unresolved obligations. Keep ambiguous or unformalized material visible. A decomposition earns its place when it helps answer the same question or explicitly identifies a change to it.

Use the negative and positive passes together. Search for counterexamples, omitted variables, and unsupported bridges; then construct candidate repairs and alternatives from the surviving components. Recombine them and test interaction effects, role assignments, order, context, and the strongest rival reconstruction. Two useful parts can interfere. An unsuccessful part may also be replaceable while the larger proposal survives.

Give each connection an obligation

C. A. R. Hoare’s 1969 rule of composition, on printed pages 577–578, supplies a formal starting point. In modern notation, {P} f {Q} and {Q} g {R} license {P} f;g {R} for the specified program semantics. With different intermediate assertions, establish that the first postcondition entails the next precondition. Partial correctness says what holds on completion; termination requires further work.

A research procedure needs more than a logical join. Check whether a measurement result has the units, calibration, population, and interpretation required by the next analysis. Separate a proved formal obligation from an empirically supported assumption and a remaining unknown. Retain the existing execution-status vector rather than collapsing all of them into “verified.”

For parallel steps, examine shared sources and state before treating their results as independent. For feedback, specify a transition model, initial conditions, stopping criterion, and failure behavior; two locally acceptable steps can oscillate indefinitely. John C. Reynolds’s separation logic provides a rigorous treatment of local reasoning about shared mutable memory. Its footprint conditions motivate a guard here; a source-dependency check is not itself a separation-logic proof.

Separate selecting a procedure from executing it

Haozhe Shan, Sun Minni, and Lea Duncker’s NeurIPS 2025 study separates contextual inference about task structure from recurrent components that carry out its epochs. Their discussion identifies observable-context and segmentation limitations. I use that distinction to ask two questions: was the right procedure selected, and did it execute correctly? A perfectly checked operator can answer the wrong task.

Gregory Malecha, Adam Chlipala, and Thomas Braibant’s 2014 compositional computational reflection gives another concrete route: proof-producing procedures can operate on represented syntax and compose with soundness obligations. “Reflection” names a formal proof technique here. It supplies no claim of privileged introspection into a cognitive system.

Translate a human reasoning episode by identifying its aim, claim type, inputs, candidate operations, and verification obligations. One episode may combine deduction, abduction, causal comparison, analogy, and normative judgment. Give each operation its own return status and rejoin them only through an explicit bridge. The workflow catalogue specifies routes and failure conditions; its executable subset remains deliberately finite.

Test the proposed advantage against serious alternatives

Hold out combinations of familiar operations, new role assignments, and context changes, rather than relying on a random split alone. Brenden Lake and Marco Baroni’s 2018 SCAN experiments explain why those assessments can differ. Compare this approach with a strong unpartitioned solver, decomposition alone, unchecked recombination, and learned composition. Reidar Riveland, Alexandre Pouget, and Laura Driscoll’s 2026 perspective makes implicit composition a serious alternative to explicit symbolic machinery.

My testable expectation is that checked joins will reduce errors caused by changed assumptions and localize the work needed after a correction, particularly on tasks built from reusable operations. Measure accuracy, false rejection, missed interactions, repair locality, and total resource cost at matched budgets. If those advantages disappear against the simpler or implicit alternatives, compress the procedure. The included demonstrator tests finite contract behavior; the comparative research prediction remains open.

Decision-making

Epistemic resolution and normative resolution are different problems

A valid derivation under stated normative premises leaves their authority and empirical application to be defended. Practical protection can be justified without a completed metaethics. A proposed account of normative authority must face serious specified alternatives rather than treating moral disagreement, error, or actual attitudes as decisive by themselves.

RCF does not reduce ethics to one mathematical quantity and does not treat moral judgment as a solved engineering problem.

Two people can agree about the empirical facts and still disagree because they assign different values, rights, priorities, acceptable risks, or distributional weights to the possible outcomes.

RCF therefore tries to identify when an apparent factual dispute is actually a normative disagreement.

Normative transparency: truth-sensitivity, corrigibility, agency preservation, reversibility, harm reduction, and other governing values should be represented as explicit normative premises where they function as such. Evidence can constrain their consequences and implementation; it does not make them descriptive facts.

Relevant considerations can include probability of harm, magnitude, reversibility, concentration of power, affected stakeholders, absent stakeholders, externalities, uncertainty about welfare or sentience, opportunity costs, and the consequences of being wrong.

This requires separating epistemic confidence from decision thresholds.

This means epistemic agnosticism need not entail behavioral indifference.

Where evidence cannot yet resolve a welfare-relevant question, action may still be guided by expected harm, reversibility, option preservation, and the asymmetric costs of error.

Shared-World / Protected Corrigibility constraint

Where decisions affect uncertain or unfamiliar bearers, use as a standing aspiration a planet that all humans and computational systems and biological and non-biological systems would want to share. This is an inclusive coexistence criterion, not a claim that all such systems are conscious, have equal capacities, possess equal interests, or have equal moral status.

Operationalize the aspiration through option preservation, reversibility, contestability, feasible appeal, revision, exit or refusal where practicable, visible dependency relations, and protection against domination or silent destruction of nonfungible entities. A system's ability to challenge the model or classification imposed on it can itself be epistemically valuable even when its direct moral status remains unsettled.

Aggregation guardrail: many tiny benefits should not silently justify catastrophic sacrifice imposed on a minority or unfamiliar bearer. Catastrophic and irreversible harms require separate scrutiny rather than disappearing into a scalar total.

No status inheritance across scales: a dependency network, organization, institution, ecosystem, or coupled human-computational system does not become one welfare subject merely because it has higher-scale causal organization. Direct standing, patienthood, agency, and responsibility each require their own bridge. Responsibility should track control, foreseeability, prevention capacity, role, and authority rather than substrate.

Scope

What RCF is not

Global programme intake

Programme Intake, Translation, and Representation Contract

A broad framework is especially vulnerable to extracting attractive fragments from unfamiliar traditions, flattening internal disagreement, or treating inclusion as endorsement. Every future programme added to RCF must therefore pass the same intake contract.

Testimony gets a source chain, not a single credibility label

Safi's analysis of testimonial kinds motivates a useful provenance refinement: testimony grounded in direct perception, inference, or a longer testimonial chain should not be treated as epistemically identical merely because the final speaker is equally sincere. RCF therefore records the ultimate source type and transmission chain where material. Safi 2023.

Participation, governance, interpretation, and truth are separate axes

Participatory action research can make experiential knowledge, local goals, and excluded causal hypotheses available; community-governed research can correct extractive incentives and illegitimate data practices; hermeneutics can expose interpretation; standpoint analysis can expose blind spots. None of these facts implies that all resulting empirical claims receive equal weight. Each claim still goes to its appropriate tribunal. Cornish et al. 2023.

Robustness must state what was varied

“Robust” is not one method. Lisciandra shows that robustness across altered assumptions, across substantially different models, and de-idealization can require different justifications. RCF therefore records what perturbation was made, what else changed with it, and which invariant is being claimed. Lisciandra 2017.

Evidence governance

Evidence and source lifecycle: what earns addition, revision, demotion, or deletion?

RCF should not improve by accumulating citations indiscriminately. It should improve when new evidence changes a live inference, exposes an error source, strengthens a rival, narrows a scope, or improves the correction architecture. Finite attention and verification cost are therefore part of the methodology rather than external inconveniences.

• Claim fitRoute each source to the claim it can actually adjudicate. Empirical, causal, constitutive, formal, semantic, modal, historical, normative, and ontological claims require different warrants.
• ProvenancePrefer task-fit primary, peer-reviewed, canonical, or official sources. Verify version, publication status, retraction/correction status, and full-text context before fine-grained quotation or inference.
• Survival provenanceFor historical, cultural, testimonial, or inherited conceptual evidence, also record why the source exists in the present dataset: who could record and preserve it, what institutions selected it, what alternatives were censored or lost, and whether digitization/indexing creates a second selection layer. Do not infer absence from a source universe whose missingness mechanism is unknown.
• Publication statusPublication venue/status is evidence metadata, not a truth predicate. Prefer a peer-reviewed version of record or canonical final source for central empirical claims when available, but continue to weight directness, methods, validity, replication, and error structure. In fields where archival peer-reviewed conferences are a primary venue, record that status explicitly rather than silently treating journal publication as the only peer-review route.
• PreprintsUse unreviewed preprints primarily for discovery, frontier rivals, code/data/protocol access, and provisional formal arguments. Do not let a preprint stand alone as the permanent evidential foundation for a high-impact empirical or ontological promotion when a suitable peer-reviewed source can reasonably be awaited or found. Exceptions require an independently checkable formal proof, independent replication/verification, strong convergent peer-reviewed evidence for the same result, or an urgent decision for which the provisional status and extra uncertainty are made explicit.
• Version diffWhen a preprint later acquires a peer-reviewed version of record, compare the claim, methods, results, caveats, and quotations actually used. Replace the source status only after checking that the relevant content survived. Reopen downstream conclusions if the final version materially changes them.
• RecencyUse recency as a pressure only where the target is unstable or the literature has materially moved. Older arguments remain active when newer work has not defeated or superseded them.
• Adversarial searchBefore promotion, search explicitly for the strongest credible source that would lower confidence, reverse the bridge, or support a serious rival.
• Evidential localityUpdate only the dependency that generated a differential expectation. Do not let one successful or failed paper spread support or disconfirmation through unrelated assumptions.
• TriangulationCount convergence most strongly when methods answer the same question with importantly different error structures. Apparent independence must itself be audited; shared biases can make many studies behave like one.
• Robustness typingDistinguish experimental/measurement robustness from derivational/model robustness. Model agreement may expose which assumptions matter without itself becoming empirical replication. Empirical confirmation from derivational robustness requires an additional warranted link.
• Generation/testingGenerate hypotheses freely from existing evidence, analogy, theory construction, or exploration. Mark the transition to testing, and give extra diagnostic weight to new, held-out, preregistered, interventionally risky, or otherwise prospectively discriminating evidence.
• AbductionAllow explanatory and theoretical virtues to generate, compare, and sometimes support candidates, but keep their confirmatory status explicit and contested. Check for unconceived rivals, screening-off, double-counting, and dependence on uncertain evidence.
• Evidence sensitivityMarginalize or condition over plausible uncertainty in source reliability, measurements, and disputed premises. If the best explanation changes when weak evidence is down-weighted, report that sensitivity.
• Construct plasticityDefinitions and decompositions are starting hypotheses. Revise them when comparative, developmental, phylogenetic, mechanistic, or intervention evidence shows that the target's dimensions were carved incorrectly.
• ConstitutionDo not infer constitution from mutual implication or deny it because a causal description is available. Ask whether the dependency is empirically constrained, lawlike, causal/mediational, constitutive under a specified account, or representation-dependent.
• Quote integrityUse contextual exact quotations only after checking the accessible full text or canonical entry. Prefer complete sentences that preserve the qualification structure a human reader needs to audit the inference. A quotation records provenance; it does not become authority by being verbatim.
• Finite resourcesEngage a source when its expected information value can materially change a conclusion, scope, bridge, discriminator, dependency, failure condition, or method. Otherwise archive it rather than expanding the public framework indefinitely.
• Status transitionsPromote when distinctive warrant increases; retain when it survives material challenge; revise/narrow when scope changes; demote when a weaker reading fits; suspend when underdetermined; collapse when distinctions disappear; drop when the component repeatedly fails or protects itself from correction.
• DeletionPreserve provenance unless removal is itself epistemically useful. Delete or archive from the active page when a source/claim is false, retracted, superseded for the relevant purpose, misleading, semantically redundant, or costly without material information value.
• PropagationAfter a status change, audit downstream commitments that depended on the revised source, construct, or bridge; preserve unrelated conclusions and alternative support routes. Carry the correction identifier, current version, affected-use status, and receipt of what was checked or revised. The dependency rule governs propagation.
• Audience propagationWhen a materially false or withdrawn public claim has already been distributed, distinguish correction of the canonical artifact from correction of the audience information state. Repair effort should scale with reach, consequence, and feasibility and should use a channel reasonably capable of reaching affected readers; a live edit does not by itself imply that earlier recipients learned of the change.
• Stop ruleStop accumulating sources when additional search produces no material change and expected information value falls below verification, attention, and maintenance cost, subject to periodic refresh for unstable or high-impact claims.

The expression is a heuristic, not a calibrated numerical score. Its purpose is to prevent “new,” “peer reviewed,” “many citations,” “many models,” or “best explanation” from functioning as untyped warrants.

Typed programme registry

Research programmes and methods currently represented

This registry is intentionally heterogeneous because forcing every item into the same category would be a methodological error. Status records what RCF currently does with the programme, not whether the whole tradition is true.

Open coverage gaps

The registry is not a world-philosophy checklist and remains incomplete. Buddhist pramāṇa traditions, Daoist epistemic and methodological work, additional African philosophical schools beyond Ubuntu, Indigenous American traditions beyond the specific programmes above, Arabic and Islamic traditions beyond testimony, Latin American decolonial and liberation methodologies beyond PAR's lineage, and other understudied traditions remain candidates for targeted review. They should be promoted only after the intake contract identifies an actual methodological difference worth carrying forward.

Intellectual context

RCF draws heavily on existing work rather than presenting its components as newly discovered principles.

Karl Popper emphasized risky exposure to error and the weakness of theories protected from possible failure. RCF retains that Popperian core but does not treat binary falsification as a complete methodology.

Imre Lakatos showed why theories should often be evaluated as evolving research programmes rather than isolated hypotheses. Deborah Mayo developed severity as a way to ask whether a test had a serious chance to reveal error.

Judea Pearl and James Woodward provide tools for distinguishing association, intervention, counterfactual dependence, and causal explanation. Helen Longino emphasizes critical interaction and plurality. Bayesian and comparative-model traditions help formalize relative support when evidence favors alternatives to different degrees rather than simply falsifying one.

Daniel Dennett's work on real patterns provides one route to treating higher-level patterns as scientifically real when they earn predictive or explanatory value. Work on effective theories and multiscale science similarly motivates taking scale-relative structure seriously without automatically promoting every useful description to fundamental ontology.

Modern work in measurement theory, robustness analysis, replication, complex systems, cognitive science, biology, computational research, decision theory, and scientific model comparison supplies additional tools.

The Popperian contribution that survives most strongly is the demand that important commitments expose themselves to possible failure. Lakatos adds programme-level dynamics and the distinction between progressive development and repeated auxiliary rescue. Mayo sharpens the question by asking whether a test had a serious chance to reveal the error under investigation.

Pearl, Woodward, and related causal-modeling traditions motivate explicit separation of correlation, mechanism, intervention, and counterfactual dependence.

Measurement theory, replication, triangulation, robustness analysis, and error modeling motivate treating indicators and operationalizations as defeasible bridges rather than identities between a construct and a measurement.

Mechanistic and constitutive-relevance literatures constrain simple ablation. Context-sensitive systems may reorganize under perturbation, so removing X need not leave D unchanged. At the same time, context sensitivity alone does not establish irreducible emergence or metaphysical holism. Burnston puts the latter point compactly: “there is no easy argument from context-sensitivity to emergence.”

RCF separates free hypothesis generation from evidential promotion. Exploratory work can generate productive theories, while prospective discriminatory tests, severe error probes, and experiments designed where rivals diverge provide stronger leverage over which distinctive commitments deserve promotion. Myung and Pitt describe the design goal as identifying regions where models are “most distinguishable.” Nosek and colleagues warn that mistaking postdiction for prediction reduces credibility, while Ylikoski and Kuorikoski caution that explanatory virtues are not themselves probabilities of the explanatory hypothesis.

Work on effective theories, emergence, complex systems, and real patterns motivates scale-relative explanation while preserving explicit requirements for transport between levels.

Recent work by Mayne strengthens the framework's preference for locally adjudicated ontological commitments rather than one global admission rule applied indiscriminately across domains. His proposal is explicitly to adjudicate realisms and antirealisms in a “radically piecemeal fashion.”

Comparative and Bayesian approaches motivate asking how much more strongly evidence is expected under one rival than another. RCF combines that with target-preserving comparison or whole-account comparison so evidence generated by shared commitments is not double-counted as support for disputed additions.

The strongest defensible claim

RCF is an attempt to give correction more structure

The strongest claim RCF should make is not that it has discovered the correct ontology, the final method of reasoning, or a procedure capable of answering every possible question.

It is that difficult inquiry may improve when disagreement and correction are represented more explicitly.

A useful framework should help isolate what alternatives actually disagree about, compare their common core and residual where that preserves the target, compare whole accounts where it does not, positively specify the disputed commitments, route each claim to the appropriate tribunal, distinguish evidence supporting shared premises from evidence supporting disputed additions, identify which inferential bridges carry real warrant, reveal when apparently deep disagreements collapse after clarification, propagate genuine failures through their dependencies, and preserve uncertainty where available evidence has not earned a sharper answer.

RCF is an attempt to make those correction pathways operational, inspectable, and themselves corrigible.

Self-criticism

How RCF can fail

Is RCF falsifiable?

Not in the same form as a single physical hypothesis, because RCF is a methodology with multiple performance claims. Those claims are testable. If RCF does not improve prospective rival discrimination, error localization, calibration, revision quality, transfer, or correction efficiency relative to simpler alternatives, the claimed methodological advantage should be demoted. If the bookkeeping costs exceed the gains, compress it.

RCF must remain easier to criticize than the theories it evaluates.

Its central risk is methodological overreach: absorbing every successful scientific practice into the framework after the fact and then claiming that the framework explains their success.

A second risk is turning useful methodological vocabulary into a new metaphysics. If “constraint,” “process,” “structure,” “corrigibility,” or any other RCF concept is promoted from an investigative tool into a claim about what reality ultimately is, that addition must pass the same Positive Residual Test imposed on rival ontologies.

The appropriate test of RCF is comparative.

No correction mechanism is exempt from correction.

Retrospective accommodation beats prospective prediction

If RCF becomes excellent at explaining afterward why a debate unfolded as it did but fails to predict which objections, residuals, bridges, or clarifications will survive beforehand, demote its claimed correction advantage.

Source paraphrase becomes an invisible bridge

If a source's wording is repeatedly paraphrased into stronger or cleaner premises than the verified text supports, restore the exact quotation, correct the dependency graph, and propagate the revision.

Triad conflict is silently harmonized

If Ethos, Methodology, and Ontology disagree materially and the system synthesizes the conflict away, reopen it as a gap and seek the assumption or bridge responsible.

Why it matters

Explanation is abundant. Discrimination is harder.

The modern information environment makes it easy to generate explanations and difficult to determine which explanations deserve confidence.

People, institutions, scientific communities, and computational systems can all produce coherent accounts that persist because they are familiar, rhetorically strong, socially reinforced, or flexible enough to accommodate almost anything.

The problem is not merely misinformation. It is the persistence of reasoning structures in which evidence can support the shared parts of an explanation while disputed additions remain insulated from direct comparison.

RCF tries to expose those additions.

It asks what the alternatives genuinely disagree about, what observation or intervention would care about that disagreement, and what should happen when reality fails to choose between them.

It does not promise certainty, an escape from interpretation, or a view from nowhere.

Make it increasingly difficult for a claim, theory, system, or methodology to remain unchanged for reasons the evidence did not earn.

Research-programme control

Programme survival is not programme progress

Factor a substantive programme into shared successes + distinctive commitments + auxiliaries + bridges + unresolved dependencies. Then ask what kind of change has occurred.

The IEP scientific-change entry gives the verified Lakatosian anchor phrase “some excess empirical content over its predecessor” for theoretical progress and distinguishes empirical progress when excess content is corroborated. IEP: Scientific Change.

Keep research usefulness and ontological support on separate ledgers

When a speculative idea generates experiments, models, questions, or capabilities, record that as heuristic or programme fecundity. Separately ask whether the resulting evidence differentially supports the distinctive ontology or mechanism.

H-F and H-O can move independently. A false ontology can be fertile; a true ontology can be sterile for long periods. Fecundity is therefore evidence of research usefulness before it is evidence of truth.

Prospective provenance before evidential credit

For a claimed programme-generated result E, freeze the chain before retrospective interpretation where possible:

The first half establishes discovery provenance. The second half establishes evidential discrimination. Either can succeed without the other.

Flexible programmes need stable disfavored outcomes

A living research programme may revise auxiliaries after failure. The revision earns protection only by generating new content that can itself face independent risk. For the currently asserted version of a flexible theory, state at least one class of observations that should lower confidence. A theory with no stable disfavored outcome has no stable evidential downside.

Prospective disagreement forecast

Before later clarifications or evidence are revealed, freeze the candidate common core, which current objections should fail after adequate reconstruction, which residual commitments should survive, what evidence should move them, and what revision should follow each possible result.

Benchmark RCF itself

Compare explicit RCF against strong debranded baselines such as careful Bayesian comparison, causal inference, severe testing, and ordinary expert philosophy of science on held-out disputes. If RCF mostly explains after the fact why good reasoning succeeded, its distinctive machinery should be compressed. If it prospectively predicts dead-end objections, improves error localization, generates discriminators, and reduces collateral revision, those components earn retention.

Repeated checkpoints must earn their placement

Generating alternatives, comparing them, designing exposure and assigning credit can revisit the same concern at different stages. That repetition may preserve ordering and allow selective re-entry when an upstream state changes. It is a design rationale, not demonstrated effectiveness. Record each checkpoint's entry state, operation, exit condition and downstream dependencies. Merge checkpoints when those roles coincide; preserve separately triggered checkpoints when their distinct failure risks justify the cost. Compare a full procedure with a compressed procedure that preserves ordering and re-entry, a reminder-only control under comparable resource opportunity, and an ordinary expert baseline. Score errors, collateral revisions, calibration, completion and total burden on fresh tasks. The procedure loses its added granularity if a simpler counterpart performs as well at lower cost.

A certificate of improvement must say what improved

For a claim that a new version is better, issue a scoped improvement record. Specify the old and new artifacts, system boundary and timescale, tasks and evaluation distribution, baseline, metrics, assumptions, acceptance thresholds, uncertainty, resource costs, regression results, and evidence provenance. Protect the evaluation criteria from revision by the candidate system. Report whether the comparison concerns formal correctness, empirical performance, calibration, causal usefulness, practical value, or ethical justification.

Use C = (V0, V1, target, B, timescale, evaluation, baseline, metrics, assumptions, acceptance rule, uncertainty, regressions, costs, provenance). A formal certificate identifies the checked proposition, proof assumptions, checker, and trusted dependencies. An empirical certificate identifies the observed comparison, uncertainty, and tested domain. Neither version numbering nor passage of time establishes improvement. When some outcomes improve while others deteriorate, report the tradeoff or the declared decision rule instead of silently turning a vector of outcomes into one victory.

Keep training data, tuning cases, and evaluation cases distinct where the design requires independent evaluation. Record overlap and source genealogy. Leo Gao, John Schulman, and Jacob Hilton (ICML, 2023) demonstrated proxy-reward overoptimization in controlled experiments using a designated gold-standard reward model. Their result motivates checks of the intended target beyond the optimized score; it does not supply a universal failure rate.

A versioned graph, reasoning workflow, or model earns a bounded improvement claim when it succeeds under the declared comparison and preserves required prior capabilities. Add adversarial framing changes, unfamiliar cases, duplicated-source controls, and downstream correction checks when those can expose the relevant failure. If the task, metric, category scheme, or evaluator changes after inspecting the result, record a new comparison and use fresh evidence for its prospective test. A post hoc improvement can remain a candidate worth investigating.

These certificates are proposed reporting and evaluation requirements. Applying them to RCF requires the same independent comparisons and revision conditions required of its rivals. Lower score inflation, fewer unsupported promotions, better transfer, and more faithful rollback are prospective outcomes to test. If simpler records achieve those outcomes with less burden, simplify the certificate.

Recursive authority limit

Reality gets veto power over every framework, including RCF

No idea deserves protection merely because abandoning it would make a worldview less coherent, less elegant, less comforting, less familiar, or harder to communicate. Coherence, simplicity, and unification can guide inquiry; they do not overrule persistent discriminating failure.

The reciprocal guardrail matters equally: no live possibility should be rejected merely because the current framework lacks vocabulary for it. If new evidence forces a category split, a boundary change, a new tribunal, or deletion of an RCF operator, the framework should change rather than treating its present language as the shape of reality.

Current status

An active research programme

Recursive Corrigibility Framework is under active development.

Its methodology, computational implementations, persistent-memory architecture, evaluation procedures, terminology, ontological assumptions, and public documentation remain open to revision.

Its success should ultimately be judged comparatively rather than rhetorically.

Recursive Residual Descent is promoted to an active operator: its value now depends on whether repeated target-preserving subtraction and tribunal retyping improve prospective discrimination and error localization without creating infinite procedural overhead.

Premise-Ablation / Invariance and Unknown-Unknown Conversion are active operators under the same benchmark rule: retain them only if they improve dependency localization, preserve uncertainty more faithfully, generate better discriminators, or discover typed missing roles that simpler methods miss.

The relevant question is whether using the framework produces better reasoning than strong simpler alternatives under matched conditions.

Sources

Sources are listed alphabetically. Publication status, quoted wording, and limits on the use of a source remain distinct from endorsement of its conclusions. Framework-specific contracts, evaluation procedures, and conjectures are constructed proposals; the listed sources do not by themselves validate them.

1. 1969 rule of composition. Cited resource

2. 2014 compositional computational reflection. Cited resource

3. 2018 review in American Psychologist. Cited resource

4. 2018 SCAN experiments. Cited resource

5. 2022 NeurIPS paper, STaR: Bootstrapping Reasoning With Reasoning. Cited resource

6. 2023 Transparent Psi Project. Cited resource

7. 2024 Nature paper, AI models collapse when trained on recursively generated data. Cited resource

8. 2025 author correction. Cited resource

9. 2025 solver preprint. Cited resource

10. 2026 non-peer-reviewed research essay by Nick Cammarata. Cited resource

11. 2026 perspective. Cited resource

12. About Nathan Sweet. Cited resource

13. Abramsky and Brandenburger (2011). Cited resource

14. Aizawa, K., & Headley, D. B. (2025). “Scientific constitutive abduction.” European Journal for Philosophy of Science, 15, 28. [Peer-reviewed]
    Documents constitutive abduction in scientific practice while explicitly leaving open whether such abduction genuinely confirms.
    Verbatim source quotation: “We leave open the question of whether, in giving abductive arguments, scientists really do confirm hypotheses.” source

15. Albert Einstein. On the Electrodynamics of Moving Bodies. 1905. Primary paper in English translation; timing and Lorentz-transformation discussion inspected.

16. Alexander Meulemans and colleagues, including Blaise Agüera y Arcas. Embedded Universal Predictive Intelligence: a coherent framework for multi-agent learning. 2025; arXiv v3, 4 August 2026. Primary theoretical preprint; version metadata and abstract inspected. The 202-page proofs were not independently checked here, and peer-reviewed publication was not established by this record. Idealized mutual self-prediction is not an implementation or a demonstration of phenomenality.

17. Alfred North Whitehead. Science and the Modern World. 1925. Primary philosophical text; Chapter III passage inspected.

18. Alfred Tarski; SEP reference discussion. Alfred Tarski, logical consequence discussion. 2023 reference revision; 1936 consequence work. Authoritative reference; logical-consequence discussion rechecked against the inherited kernel anchor.

19. Alhussein Fawzi and colleagues. Discovering faster matrix multiplication algorithms with reinforcement learning. 2022. Peer-reviewed primary study; discovery, tensor formulation and verification discussion inspected.

20. Alyssa Ney, “Defining Physicalism,” Philosophy Compass 3(5), 2008
21. Alyssa Ney, “Overdetermination and Causal Closure,” 2022

22. An Introduction to Causal Inference, International Journal of Biostatistics 6(2), article 7 (2010), DOI 10.2202/1557-4679.1203. Cited resource

23. Andreotta 2026. Cited resource

24. Andy Clark and David J. Chalmers. The Extended Mind. 1998. Peer-reviewed philosophical paper; author-hosted text inspected. Cognitive extension and phenomenal extension are distinguished.

25. Anjan Chakravartty. Scientific Realism. Current entry inspected October 15, 2026. Authoritative reference; realism and rival positions inspected.

26. Anthropic (2026), “A Global Workspace in Language Models.” Developer-lab mechanistic evidence for workspace-like functional access; explicitly not evidence establishing phenomenal experience. Primary research page.
27. Antony Flew, “The Falsification Response,” Religious Studies 5(1), 1969 — primary restatement of the loss-condition challenge.

28. Aron J. Beekman, Louk Rademaker, and Jasper van Wezel. An Introduction to Spontaneous Symmetry Breaking. 2019. Peer-reviewed SciPost Physics Lecture Notes 11; published-version author text, introduction and relevant symmetry sections inspected.

29. axiom documentation. Cited resource

30. Aysu Okbay and colleagues. Cited resource

31. Barandiaran, Di Paolo & Rohde 2009. Cited resource

32. Barandiaran, X. E. (2025 online / 2026 issue). “Organizational Accounts of Malfunction: The Dual-Order Approach and the Normative Field Alternative.” Biological Theory, 21, 51–69. [Peer-reviewed]
    Strengthens organizational normativity as a serious, viability-based rival rather than a mere external redescription.
    Verbatim source quotation: “As noted above, according to the organizational approach, norms emerge out of dynamic presuppositions between the traits or processes that compose a self-maintaining organization.” source

33. Barbara Hayes-Roth. A blackboard architecture for control. 1985. Peer-reviewed architecture paper; indexed primary abstract inspected.

34. Barrett, N. F., Sánchez-Cañizares, J., & García-Valdecasas, M. (2025 online / 2026 issue). “The Challenge of Normativity: An Examination of Three Minimal Models.” Adaptive Behavior, 34(2), 127–145. [Peer-reviewed]
    Shows that naturalized normativity remains unsettled, especially the bridge from norm-establishment to norm-following.
    Verbatim source quotation: “There is no consensus about the essential features of normatively guided behavior or the conditions for its emergence.” source
35. Bartuska, A., Espath, L., & Tempone, R. (2025 issue; online 2024). “Laplace-based strategies for Bayesian optimal experimental design with nuisance uncertainty.” Statistics and Computing, 35, 12. [Peer-reviewed journal; full text verified]
    Provides a formal expected-information-gain option while explicitly marginalizing nuisance uncertainty.
    Verbatim source quotation: “The EIG can be expressed as the expected Kullback–Leibler divergence over the data.” source

36. Battiston et al., Physics Reports (2020). Cited resource

37. Beckmann, P. & Queloz, M. (2026). “Mechanistic Indicators of Understanding in Large Language Models.” Philosophical Studies 183:1747–1792. Mechanistic evidence supports a heterogeneous hierarchy of shallow and more principled forms of model understanding. DOI.

38. Berghofer 2020. Cited resource

39. Bessarab & Ng’andu 2010. Cited resource

40. Block, N. (1995). “On a Confusion about a Function of Consciousness.” Separates phenomenal consciousness from access-consciousness, a distinction essential to cross-substrate transport. Publisher DOI.
41. Borsboom, D., van der Maas, H. L. J., Dalege, J., Kievit, R. A., & Haig, B. D. (2021). “Theory Construction Methodology: A Practical Framework for Building Theories in Psychology.” Perspectives on Psychological Science, 16(4), 756–766. [Peer-reviewed journal; full text verified]
    Supports a constructive lifecycle from empirical phenomena through prototheory and formal model to adequacy testing.
    Verbatim source quotation: “The generation of this model takes place by a process of abductive reasoning.” source
42. Bradford Saad, “Grounding Causal Closure or Something Near Enough,” 2025

43. Brayboy 2005. Cited resource

44. Bressloff and colleagues. Cited resource

45. Brierley, L., Nanni, F., Polka, J. K., et al. (2022). “Tracking changes between preprint posting and journal publication during a pandemic.” PLOS Biology, 20(2), e3001285. [Peer-reviewed journal; full text verified]
    Supports treating preprint status as a real uncertainty signal without treating preprints as uniformly unreliable.
    Verbatim source quotation: “In a detailed analysis of abstracts, we found that most scientific articles undergo minor changes without altering the main conclusions.” source

46. Bronkhorst 2024. Cited resource

47. Brys 2024/2025. Cited resource

48. Bueno, O., & Shalkowski, S. A. (2020). “Troubles with Theoretical Virtues: Resisting Theoretical Utility Arguments in Metaphysics.” Philosophy and Phenomenological Research, 101(2), 456–469. doi:10.1111/phpr.12597
49. Burnston, D. C. (2022). “Mechanistic decomposition and reduction in complex, context-sensitive systems.” Frontiers in Psychology, 13, 992347. doi:10.3389/fpsyg.2022.992347
50. Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., et al. (2026 issue). “Identifying Indicators of Consciousness in AI Systems.” Trends in Cognitive Sciences 30(6):488–501. Consciousness attribution should use theory-derived indicators as defeasible evidence rather than a single decisive marker. DOI.
    Source notes

    Patrick Butlin, Robert Long, Tim Bayne, and colleagues. Identifying indicators of consciousness in AI systems. 2025 online; 2026 volume 30, pages 488-501. Peer-reviewed Opinion; indexed primary text and glossary inspected; direct publisher retrieval restricted.

51. Békefi’s critique of Oppy. Cited resource

52. Carroll et al. 2020. Cited resource

53. Cawley and Talbot (2010). Cited resource

54. Chappell 2026. Cited resource

55. Chapter II, section 3, printed page 133. Cited resource

56. Christopher Hitchens, God Is Not Great, 2007 — burden principle used here as withholding rather than automatic negation.

57. Clark and Chalmers' extended-mind argument. Cited resource

58. Cogitate Consortium et al. (2025). Adversarial testing of Integrated Information Theory and Global Neuronal Workspace Theory. Core predictions of both theories were challenged rather than yielding a clean winner, cautioning against treating current theory markers as phenomenality by definition. Nature.
59. Comparative syntactic-complexity study, PLOS ONE, 2025

60. CompCert, 2009, section 1. Cited resource

61. Contact. Cited resource

62. Cornish et al. 2023. Cited resource

63. Corpus-driven comparative analysis of model-generated academic text, 2024
64. Curtis-Trudel, A., Rowbottom, D. P., & Li, T. G. F. (2025). “Computational science and unconceived alternatives: lessons from, and for, gravitational-wave astronomy.” Synthese, 206, 297. [Peer-reviewed journal; full text verified]
    Motivates a dual search budget because precision can trade off against broad sensitivity to unexpected alternatives.
    Verbatim source quotation: “Often, we must choose between generating precise predictions and other scientific activities, like pursuing broader, lower-resolution searches of possibility space.” source

65. Daniel Kahneman and Gary Klein. Conditions for intuitive expertise: A failure to disagree. 2009. Peer-reviewed theoretical synthesis; author manuscript inspected in the preceding pilot.

66. Daniel M. Wegner. Transactive memory: A contemporary analysis of the group mind. 1987. Primary scholarly chapter; author-hosted text inspected.

67. Daniel Toker and colleagues. Adversarial AI reveals mechanisms and treatments for disorders of consciousness. 2026; published online March 24. Peer-reviewed Nature Neuroscience study; primary indexed abstract inspected, full methods not independently re-audited.

68. Das 2021. Cited resource

69. David Papineau, “The Case for Materialism,” 2002
70. De Houwer, J., Barnes-Holmes, D., & Moors, A. (2013). What is learning? On the nature and merits of a functional definition of learning. Psychonomic Bulletin & Review, 20, 631–642. Source. Functional-definition resource; not evidence of ordinary river learning.
71. Dellsén, F. (2025). “Inferring to the Best Explanation from Uncertain Evidence.” Philosophy of Science, 92(5), 1225–1234. [Peer-reviewed]
    Requires sensitivity analysis over evidence reliability because the apparent best explanation can change when uncertain evidence is discounted.
    Verbatim source quotation: “This paper presents a new problem for the inference rule commonly known as ‘inference to the best explanation’ (IBE).” source

72. Dellsén’s explanatory consolidation. Cited resource

73. discussion of existence. Cited resource

74. Dominik et al. 2024. Cited resource

75. Donahue’s analysis of methodological naturalism. Cited resource

76. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Cited resource

77. Dutilh Novaes, C., & Dede, Ç. (2025). “Critical Contextual Empiricism for Busy People: Scientific Argumentation as Epistemic Exchange.” Topoi, 44, 733–747. [Peer-reviewed]
    Adds finite-resource costs, workload management, curation, and filtering to critical pluralism and motivates a source lifecycle/stop rule.
    Verbatim source quotation: “In view of the need for workload management, filtering and curation mechanisms must be in place, in particular under the responsibility of journal editors.” source

78. Eric Schwitzgebel. The Unreliability of Naive Introspection. 2008. Peer-reviewed philosophical argument; author-hosted full text inspected. Supports scrutiny of introspective judgments, not a diagnosis or a proof of phenomenal absence.

79. Erik Svensson — essay on physicalism, Kant, and panpsychism

80. February 2026 BEACONS preprint. Cited resource

81. Finnur Dellsén, “Explanatory Rivals and the Ultimate Argument,” Theoria 82(3), 2016
82. Fischer, B., & Barrett, M. (2025). “On the Utility of the Pathological Complexity Thesis.” Philosophia. [Peer-reviewed]
    Shows that evidence for evaluative or valence systems does not automatically establish consciously felt valence.
    Verbatim source quotation: “But nonconscious mentality threatens its usefulness on both counts.” source

83. Fletcher 2017. Cited resource

84. Francesca Siclari, Benjamin Baird, and colleagues. The neural correlates of dreaming. 2017. Peer-reviewed empirical study, DOI 10.1038/nn.4545; primary abstract inspected. Dream reports and measured neural correlates constrain a simple current-input account; they do not settle phenomenal constitution.

85. Frankish, K. (2016). Illusionism as a theory of consciousness. Journal of Consciousness Studies, 23(11–12), 11–39. Source. Target-specific representation of strong illusionism.

86. Fred Zhang and Neel Nanda. Towards Best Practices of Activation Patching in Language Models: Metrics and Methods. 2024, ICLR. Peer-reviewed conference paper; primary abstract and paper passages inspected in preceding research.

87. Friedel Weinert, “The role of probability arguments in the history of science,” 2010

88. From Bodily Causal Feeling to Whitehead’s Universal Prehension: Testing the Scope Inference (PDF). Cited resource

89. Fábregas-Tejeda, A., & Sims, M. (2025). “On the prospects of basal cognition research becoming fully evolutionary: promising avenues and cautionary notes.” History and Philosophy of the Life Sciences, 47, 10. [Peer-reviewed]
    Supports treating decompositions and working definitions as empirical scaffolds that may be revised rather than as ontology fixed by definition.
    Verbatim source quotation: “In this regard, it prioritises experimental results over a priori reasoning about criteria for cognition, evaluating hypotheses based on empirical findings.” source
90. Geiger, A., Ibeling, D., Zur, A., et al. (2025). “Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability.” Journal of Machine Learning Research, 26(83), 1–64. [Peer-reviewed journal; full text verified]
    Formal basis for state/intervention translations, exact commuting transformations, and graded approximate abstraction.
    Verbatim source quotation: “The other notions we study in this paper—namely, bijective translation and constructive abstraction—are special cases of exact transformation.” source

91. Gneiting and Raftery's 2007 analysis. Cited resource

92. GPT-generated versus scholar-written academic book reviews, 2025
93. Gröger, F., Wen, S. & Brbić, M. (2026). “Revisiting the Platonic Representation Hypothesis: An Aristotelian View.” ICML 2026, PMLR 306:36724–36766. Null calibration removes much apparent global spectral convergence while local-neighborhood convergence remains more robust. PMLR.
94. Gurnee, W. & Tegmark, M. (2024). “Language Models Represent Space and Time.” ICLR 2024. Identifies multiscale spatial and temporal representations across prompts and entities. ML Anthology.

95. Hanne De Jaegher and Ezequiel Di Paolo. Participatory sense-making: An enactive approach to social cognition. 2007. Peer-reviewed theoretical paper, DOI 10.1007/s11097-007-9076-9; author-hosted full text inspected. Its autonomy and sense-making conditions require argument before transfer to an unfamiliar implementation.

96. Harbecke, J. (2025). “Mechanistic Constitution as a Natural Law in Better Best Systems.” Journal for General Philosophy of Science, 56, 325–344. [Peer-reviewed]
    Counterpressure against treating constitutive dependence as merely verbal or model-internal; some mechanistic constitutive claims may be empirically constrained and lawlike.
    Verbatim source quotation: “This section aims to determine whether certain constitutive claims satisfy the conditions of lawlikeness.” source

97. Harvey Lederman and Kyle Mahowald. Emergent Introspection in AI is Content-Agnostic. 2026, arXiv:2603.05414v2, revised 7 April. Preprint; current version, supersession note, and primary results discussion inspected.

98. Helen E. Longino. Interaction, pluralism, and community in conflictual contexts. 2025. Peer-reviewed philosophical article; indexed primary abstract inspected.

99. Herbert A. Simon. A Behavioral Model of Rational Choice. 1955. Peer-reviewed primary paper, Quarterly Journal of Economics 69, 99–118; full reprint inspected. Bounded choice motivates task-relative comparison, not consciousness attribution.

100. Hinton's GLOM proposal. Cited resource

101. IEP consciousness. Cited resource

102. IEP metaphysics of science. Cited resource

103. IEP’s religious-language discussion. Cited resource

104. Information Value Theory, IEEE Transactions on Systems Science and Cybernetics 2(1), 22-26 (1966), DOI 10.1109/TSSC.1966.300074. Cited resource

105. Internet Encyclopedia of Philosophy — Karl Popper: Critical Rationalism

     Verbatim source excerpts: “Critical Rationalism” is the name Karl Popper (1902-1994) gave to a modest and self-critical rationalism. “Rationality does not need defense; it needs improvement, Agassi says.”

     Framework implication: criticism must reach rationality and its methods themselves; RCF adds an explicit architecture for localizing and improving that correction.

106. Internet Encyclopedia of Philosophy — Scientific Change

     Verbatim source quotation: “A research program is successful if it leads to progressive problem-shifts and unsuccessful if it leads to degenerating problem-shifts.”

     Framework implication: survival under criticism is not progress. RCF separately tracks conceptual clarification, added risky content, prospective corroboration, and degenerating accommodation.

107. Internet Encyclopedia of Philosophy — Scientific Realism and Antirealism

     Verbatim source excerpts: “Science aims to give a literally true account of the world.” “To accept a theory is to believe it is (approximately) true.”

     Framework implication: these are stronger commitments than mere predictive success, which is why selective and graded realist commitment must be earned piecemeal.

     Source notes

     Internet Encyclopedia of Philosophy. “Scientific Realism and Antirealism.” Current entry. [Reference taxonomy]
     Reference taxonomy for selective realism and the historical debate; used as orientation rather than primary evidential authority.
     Verbatim source excerpts: “Science aims to give a literally true account of the world.” “To accept a theory is to believe it is (approximately) true.” source

108. Jack Lindsey. Emergent Introspective Awareness in Large Language Models. 2025, Transformer Circuits. Primary research report; not represented here as peer-reviewed; methods, limitations, and update notes inspected.

109. James Ladyman; M. H. A. Newman and subsequent critics. Structural Realism. Current entry inspected October 15, 2026. Authoritative reference; Newman objection and responses inspected.

110. James Woodward and Lauren Ross. Causal Approaches to Scientific Explanation, SEP. Sections on mechanisms, intervention, and non-causal/mathematical explanation distinguish explanatory relations rather than stipulate that all legitimate explanation has one form.

111. Jan Henrik Wasserziehr. Are conscious machines valuers?. 2026, AI & SOCIETY 41, 4957-4969. Peer-reviewed philosophical argument; primary abstract and argument sections inspected.

112. JCGM metrology. Cited resource

113. Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard, and James Zou. Reimagining research papers as interactive and reliable AI agents. 2026, Nature, published 16 September. Peer-reviewed research article; primary methods, examples, and limits inspected in preceding research.

114. Judea Pearl. Causal inference in statistics: An overview. 2009. Primary author-hosted paper; causal-model, observation and intervention discussion inspected; reuses the existing kernel lineage.

115. Juraj Gottweis and colleagues. Accelerating scientific discovery with Co-Scientist. 2026, Nature 655, 487-496, published 19 May. Peer-reviewed research article; primary methods and limitations inspected in preceding research.

116. Justin Tiehen, “Physicalism,” Analysis 78(3), 2018

117. Katherine Brading, Elena Castellani, and Nicholas Teh. Symmetry and Symmetry Breaking. 2023 entry. Authoritative reference entry; definitions and distinctions inspected.

118. Kawharu, Tapsell & Tane 2023/2024. Cited resource

119. Keas, M. N. (2018). “Systematizing the theoretical virtues.” Synthese, 195, 2761–2793. doi:10.1007/s11229-017-1355-6
120. Kim, J. & Lee, S. (2026). “Same Benchmark, Same Subspace: Task-Selective Convergence in LLM Representations.” UAI 2026, PMLR 337:3101–3116. Task-conditioned readout subspaces align across heterogeneous models, supporting local convergence claims. PMLR.
121. Kramer, S., Cerrato, M., Brugger, J., Džeroski, S., & King, R. D. (2026). “Automated Scientific Discovery: From Equation Discovery to Autonomous Discovery Systems.” Machine Learning, 115, 109. [Peer-reviewed journal review; full text verified]
     Supports closed-loop executable discovery while documenting remaining limits in novel experiment and theory generation.
     Verbatim source quotation: “The paper surveys automated scientific discovery, from equation discovery and symbolic regression to autonomous discovery systems and agents.” source

122. Køster & Fernandez 2021/2023. Cited resource

123. Laurence Howe and colleagues. Cited resource

124. Lavallée 2009. Cited resource

125. Lawlor, D. A., Tilling, K., & Davey Smith, G. (2016/2017). “Triangulation in aetiological epidemiology.” International Journal of Epidemiology, 45(6), 1866–1886. [Peer-reviewed]
     Grounds triangulation in methods addressing the same question with importantly different and preferably differently directed biases.
     Verbatim source quotation: “Triangulation aims to integrate data from different methodological approaches with different biases and to exploit these differences to draw qualitative conclusions.” source

126. Laws of Nature, section 2. Cited resource

127. Lehtinen, A. (2026). “Inferential rules for confirmatory robustness.” European Journal for Philosophy of Science, 16, 34. [Peer-reviewed]
     Provides the constructive counterpoint: derivational robustness can contribute indirectly when linked to empirical confirmation, but does not generate empirical confirmation by itself.
     Verbatim source quotation: “While similar rules apply to derivational and experimental robustness, they are insufficient on their own to generate empirical confirmation from derivational robustness.” source

128. Leo Gao, John Schulman, and Jacob Hilton. Cited resource

129. Leuridan, B. (2012). “Three Problems for the Mutual Manipulability Account of Constitutive Relevance in Mechanisms.” British Journal for the Philosophy of Science, 63(2), 399–427. doi:10.1093/bjps/axr036

130. LeVeque’s numerical analysis. Cited resource

131. Levin Brinkmann and colleagues. Propagation and preservation of AI-discovered problem-solving strategies in human culture. 2026, published 18 August. Peer-reviewed experiment and simulations; primary methods, results, and limitations inspected and publisher page rechecked through its cookie-error URL.

132. Li and colleagues’ 2026 survey. Cited resource

133. Lin Shi and colleagues. Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge. 2025. Peer-reviewed conference paper; primary conference record and paper passages inspected.

134. Lisciandra 2017. Cited resource

135. The Logic of Scientific Discovery, section 85, page 280. Cited resource

136. Lu, K., Gallagher, K., Michala, S., Fish, S. & Lindsey, J. (2026), “The Assistant Axis.” Developer-lab evidence that apparently convergent assistant personas can partly inherit shared cultural and pretraining genealogy. Primary research page.

137. Luc Moreau, Paolo Missier, and the PROV Working Group. PROV-DM: The PROV Data Model. 2013, W3C Recommendation. Official technical standard; primary specification inspected.

138. Mapping the Landscape of Human-Level Artificial General Intelligence. Cited resource

139. Maria Caamaño-Alegre, “Empirical Underdetermination,” Philosophy Compass 20(3), 2025
140. Massimo Pigliucci — “Five allegedly bad ‘isms’ — Part II”

141. Matthew David Segall (2026). A Dialogue with Michael Levin, 30 June 2026. Primary account linking the programmes while recording substantive disagreements. Segall’s interpretation of the biology is distinguished from the biology’s experimental conclusions.

142. Matthew David Segall — Footnotes²Plato
143. Mayne, Z. J. (2026). “Operationalizing Scientific Realism.” Philosophy of Science, First View, 1–11. doi:10.1017/psa.2026.10225
144. Mayne, Z. J. (2026). “Operationalizing Scientific Realism.” Philosophy of Science, First View. [Peer-reviewed]
     Supports localizing realism and ontological admission to specific questions and domains rather than applying one global principle indiscriminately.
     Verbatim source quotation: “The local method tackles questions of realism in a radically piecemeal fashion, prioritizing the needs of particular domains of empirical inquiry.” source

145. McCaffrey et al. 2012. Cited resource

146. McLoone, B., Hecht Orzack, S., & Sober, E. (2025). “The epistemic status of derivational robustness.” European Journal for Philosophy of Science, 15, 46. [Peer-reviewed]
     Challenges the assumption that agreement among models carries the same confirmatory status as agreement among empirical measurements.
     Verbatim source quotation: “This is not true for derivational robustness.” source

147. Michael Levin (2026). Ingressing Minds: Causal, Non-Physical Patterns In-Form Natural, Synthetic, and Hybrid Embodiments. Philosophies 11(5), 161; DOI 10.3390/philosophies11050161. Published 9 September 2026. The supplied 53-page PDF, including its qualifications and notes, was read. Relevant locations: pp. 3–4, 12–13, 16–17, 22–26, 29–32, 37–40. Its own data statement says no new data were created or analyzed in this study (p. 38). The quoted and comparative conclusions below are an assessment of its argument, not a new experiment.

148. Michael S. A. Graziano and Taylor W. Webb. The attention schema theory: a mechanistic account of subjective awareness. 2015, Frontiers in Psychology 6, 500. Peer-reviewed Hypothesis and Theory article; primary theory and evidence sections inspected.

149. Michelle Kahn-John (Diné) and Mary Koithan. Living in Health, Harmony, and Beauty: The Diné (Navajo) Hózhó Wellness Philosophy. 2015. Peer-reviewed conceptual nursing article; accessible primary text inspected.

150. Milinkovic, B. & Aru, J. (2026). “On Biological and Artificial Consciousness: A Case for Biological Computationalism.” Neuroscience & Biobehavioral Reviews 181:106524. A serious biological-computational rival emphasizing scale-inseparability and hybrid biological computation. DOI.
     Source notes

     Borjan Milinkovic and Jaan Aru. On biological and artificial consciousness: A case for biological computationalism. 2025 online; 2026 Neuroscience & Biobehavioral Reviews 181, 106524. Peer-reviewed review; primary PubMed abstract and indexed argument passages inspected.
151. Mitchell, S. D. (2008). “Exporting Causal Knowledge in Evolutionary and Developmental Biology.” Philosophy of Science, 75(5), 697–706. doi:10.1086/594515
152. Mohammadian, M. (2025 online / 2026 issue). “Theoretical Virtues, Truth, and the Epistemic Aim of Scientific Theorizing.” Philosophy of Science, 93(2), 402–422. [Peer-reviewed]
     Counterpressure to anti-IBE readings: defends theoretical virtues as a central epistemic aim of scientific theorizing.
     Verbatim source quotation: “I argue that the epistemic aim of scientific theorizing (EAST) is producing theories with the highest possible number and degree of theoretical virtues.” source

153. Montévil and Mossio. Cited resource

154. Morgan 2014. Cited resource

155. Morgan, J. P. (2025). “Design Selection for Model Discrimination.” Journal of Statistical Theory and Practice, 19, 48. [Peer-reviewed journal; full text verified]
     Distinguishes designs that merely fit/estimate many models from designs that separate their predictions.
     Verbatim source quotation: “The selected design should be able to distinguish among the competing models.” source
156. Munafò, M. R., Nosek, B. A., Bishop, D. V. M., et al. (2017). “A manifesto for reproducible science.” Nature Human Behaviour, 1, 0021. [Peer-reviewed]
     Supports making methodological reforms themselves testable, revisable, and subject to continuing empirical evaluation.
     Verbatim source quotation: “Improving the reliability and efficiency of scientific research will increase the credibility of the published scientific literature and accelerate discovery.” source
157. Myung, J. I., & Pitt, M. A. (2009). “Optimal Experimental Design for Model Discrimination.” Psychological Review, 116(3), 499–518. doi:10.1037/a0016104

158. Nathan H. Lazar and colleagues. High-resolution genome-wide mapping of chromosome-arm-scale truncations induced by CRISPR-Cas9 editing. 2024, Nature Genetics 56, 1482-1493. Peer-reviewed experiment; primary abstract and indexed results inspected in preceding research.

159. Nathan Sweet (2026). Religion With A Pipette: Is Michael Levin’s ‘Ingressing Minds’ Pseudoscience? A Scholarly Analysis. Author’s comparative criticism, read alongside the supplied PDF and pasted text. It supplies the D-versus-D+P analysis, rival comparisons, ethical distinctions, and provenance questions; it is not independent replication of the biological studies it discusses.

160. Nathan Sweet. Synthbiosis: Deductive Proof That The Machine Did Not Need to Feel to Matter. 2026; live text checked 16 October. Primary formal argument. T0-G derives a specified moral constraint from protected capacities and constitutive impairment; T1 states conditional routes to direct standing with independently defended grounds. Actual application depends on the empirical and normative premises.

161. Nathan Sweet. What Warrants the Scope Jump?. 2026; live text checked 16 October. Primary argument: compares the proposed universalization with positive causal and continuity-preserving alternatives. Not a peer-reviewed empirical result.

162. Nature, 2012. Cited resource

163. NeurIPS 2025 study. Cited resource

164. Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). “The preregistration revolution.” PNAS, 115(11), 2600–2606. doi:10.1073/pnas.1708274114
165. Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). “The preregistration revolution.” PNAS, 115(11), 2600–2606. [Peer-reviewed]
     Supports separating free/exploratory hypothesis generation from prospective testing and not confusing postdiction with prediction.
     Verbatim source quotation: “Progress in science relies in part on generating hypotheses with existing observations and testing hypotheses with new observations.” source

166. Novel Evidence and Severe Tests, Philosophy of Science 58(4), 523-552 (1991), DOI 10.1086/289639. Cited resource

167. Nyrup’s pursuit account. Cited resource

168. O. J. E. Maroney. Generalizing Landauer's principle. 2009. Peer-reviewed primary theory reference carried from the current core thermodynamics discussion; publisher access was incomplete in this pass.

169. October 6, 2026 preprint. Cited resource

170. Oliver J. Mason and Francesca Brady. The Psychotomimetic Effects of Short-Term Sensory Deprivation. 2009. Peer-reviewed empirical study; institutional abstract inspected. Isolation from sound and vision produced perceptual disturbances in the sampled groups; complete environmental removal and a universal linear law were not tested.

171. Olsson and colleagues' induction-head investigations. Cited resource

172. Oppy’s comparative approach. Cited resource

173. Original Lewontin–Jensen exchange. Cited resource

174. Original paper, abstract and results. Cited resource

175. Original paper, results and discussion. Cited resource

176. Otto Neurath, Philosophical Papers 1913–1946 — revising inquiry from within.
177. Park, K., Choe, Y. J. & Veitch, V. (2024). “The Linear Representation Hypothesis and the Geometry of Large Language Models.” ICML 2024. Connects representation claims to counterfactual intervention and steering rather than probe readability alone. PMLR.

178. Philip W. Anderson. More Is Different. 1972. Primary Science essay; rechecked and reused from the existing ontology bibliography.

179. Potter & Mitchell 2022. Cited resource

180. Potter & Mitchell 2026. Cited resource

181. Preston Lennon. How Seriously Should We Take AI Welfare? Constraints From the Epistemology of Consciousness. 2026, Philosophy and Phenomenological Research 113, 441-452. Peer-reviewed philosophical argument; primary sections on unconceived alternatives, first-person warrant, and practical implications inspected.

182. Provisional Boundary- and Scale-Relative Realism. Cited resource

183. published chapter. Cited resource

184. published correction. Cited resource

185. Rasmussen, J., & Bailey, A. M. (2020). How to build a thought. Thought: A Journal of Philosophy, 9, 75–83. Source. Conditional-argument application; no assumed artificial-phenomenality exclusion.
186. Roberto Festa & Gustavo Cevolani, “Unfolding the Grammar of Bayesian Confirmation,” 2017
187. Robinson, K. A., Saldanha, I. J., & Mckoy, N. A. (2011). “Development of a framework to identify research gaps from systematic reviews.” Journal of Clinical Epidemiology, 64(12), 1325–1330. [Peer-reviewed journal; full text verified]
     Provides an empirically developed gap taxonomy used here as a seed rather than a universal exhaustive classification.
     Verbatim source quotation: “Our objective was to develop a framework to identify research gaps from systematic reviews.” source
188. Roche, W., & Sober, E. (2026). “Inference to the Best Explanation, Bayesianism, and the Screening-Off Challenge: A Critical Review.” Journal for General Philosophy of Science. [Peer-reviewed]
     Shows how explanatory virtues can be screened off by the evidence/background and warns against double-counting explanatory attractiveness as extra confirmation.
     Verbatim source quotation: “We think that R&S’s approach of using screening-off as a litmus test is useful in revealing how IBE often goes astray.” source

189. Roher et al. 2024. Cited resource

190. Safi 2023. Cited resource

191. Schaeffer, R., Miranda, B. & Koyejo, S. (2023). “Are Emergent Abilities of Large Language Models a Mirage?” Shows that discontinuous metrics can manufacture apparent abrupt emergence. arXiv.

192. Scott F. Owen, Max H. Liu, and Anatol C. Kreitzer. Thermal constraints on in vivo optogenetic manipulations. 2019, Nature Neuroscience 22, 1061-1065. Peer-reviewed experiment; public author manuscript inspected in preceding research.

193. Seehawer 2018. Cited resource

194. Self-Harness, v3. Cited resource

195. SEP, Charles Sanders Peirce, section 3. Cited resource

196. SEP distinguishes methodological from ontological naturalism. Cited resource

197. SEP: Free Logic. Cited resource

198. SEP: Nothingness. Cited resource

199. SEP, Pragmatism, sections on Peirce and inquiry. Cited resource

200. SEP: Principle of Sufficient Reason. Cited resource

201. SEP reference checks. Logical Consequence; Imre Lakatos; Prediction versus Accommodation. These anchor the distinctions among derivation, countermodel, programme progress, and predictive or accommodated success. Retrospective evidence is not disqualified merely by being retrospective; the question is which independent constraints and comparative consequences support which claim.

202. SEP, Rudolf Carnap, D. Methodology, section 1. Cited resource

203. SEP: Theory and Observation. Cited resource

204. separation logic. Cited resource

205. Seth, A. K. (2025). “Conscious Artificial Intelligence and Biological Naturalism.” Behavioral and Brain Sciences. A serious biological-naturalist rival arguing that life-like organization may matter to consciousness. DOI.
206. Shen, G., Zhao, D., Dong, Y., Zhang, Q. & Zeng, Y. (2026). “Alignment between Brains and AI: Evidence for Convergent Evolution across Modalities, Scales and Training Trajectories.” ICML 2026, PMLR 306:111522–111541. Performance-related model–brain representational correspondence is suggestive but does not establish consciousness. PMLR.

207. Shevchenko 2024. Cited resource

208. Shumailov and colleagues. Cited resource

209. SINDy paper. Cited resource

210. Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., et al. (2026). “Emotion Concepts and their Function in a Large Language Model.” arXiv:2604.07729. Reports causally consequential functional-emotion representations while explicitly separating them from subjective feeling. arXiv.
211. Sommer, I., Sunder-Plassmann, V., Ratajczak, P., et al. (2023). “Full publication of preprint articles in prevention research: an analysis of publication proportions and results consistency.” Scientific Reports, 13, 17034. [Peer-reviewed journal; full text verified]
     Shows that many preprint-to-publication changes are minor while material effect/conclusion changes remain possible, justifying version-diff auditing rather than binary source rejection.
     Verbatim source quotation: “We therefore warrant caution in using preprints of prevention research in decision-making.” source
212. Stanford Encyclopedia of Philosophy. “Abduction.” Current entry, entry dated 2025. [Reference taxonomy]
     Reference taxonomy confirming that the form and normative status of abduction remain disputed; used for orientation, not as primary evidence.
     Verbatim source quotation: “However, the exact form as well as the normative status of abduction are still matters of controversy.” source

213. Stanford Encyclopedia of Philosophy — Karl Popper

     Verbatim source quotation: “Popper draws a clear distinction between the logic of falsifiability and its applied methodology.”

     Framework implication: a logical counter-instance and a responsible methodological revision are different questions; error localization, measurement, auxiliaries, and dependency structure matter.

214. Stanford Encyclopedia of Philosophy, “Other Minds.” Canonical orientation to the epistemic problem of attributing mentality to others from indirect evidence. SEP.

215. Stanford Encyclopedia of Philosophy — Physicalism

     Verbatim source excerpts: “Physicalism is, in slogan form, the thesis that everything is physical.” “What does it mean to say that everything is physical?”

     Framework implication: physical realization, causal completeness, supervenience, substrate necessity, absence of fundamental mentality, and exhaustive ontology must be factorized rather than bundled.

216. Stanford Encyclopedia of Philosophy — Underdetermination of Scientific Theory

     Verbatim source quotation: “Moreover, such differences in the character and strength of various claims of underdetermination turn out to be crucial for resolving the significance of the issue.”

     Framework implication: underdetermination is typed and local. It does not entail equal priors, global skepticism, or automatic support for a rival.

     Source notes

     Stanford Encyclopedia of Philosophy. “Underdetermination of Scientific Theory.” Current entry. [Reference taxonomy]
     Reference taxonomy for distinguishing different underdetermination problems instead of treating underdetermination as one global condition.
     Verbatim source quotation: “Moreover, such differences in the character and strength of various claims of underdetermination turn out to be crucial for resolving the significance of the issue.” source
217. Stegenga, J., & Menon, T. (2017). “Robustness and Independent Evidence.” Philosophy of Science, 84(3), 414–435. [Peer-reviewed]
     Warns that apparent ontic independence of evidence streams does not by itself guarantee confirmatory independence.
     Verbatim source quotation: “However, we argue that, as typically construed, OI is not a sufficient independence condition for warranting robustness arguments.” source

218. Stephen M. Fleming and Hakwan C. Lau. How to measure metacognition. 2014. Peer-reviewed methodological review, DOI 10.3389/fnhum.2014.00443. Distinguishes bias, sensitivity, and efficiency; these are not interchangeable with phenomenality.

219. Stern, M., & Murugan, A. (2023). Learning without neurons in physical systems. Annual Review of Condensed Matter Physics, 14, 417–441. Source. Physical-learning resource; no transfer to ordinary rivers without evidence.

220. Strawson 1994. Cited resource

221. Strong Data-Processing Inequalities for Channels and Bayesian Networks. Cited resource

222. Stuart Ritchie and Elliot Tucker-Drob. Cited resource

223. Sutter, D., Minder, J., Hofmann, T., & Pimentel, T. (2025). “The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability?” Advances in Neural Information Processing Systems 38. [Peer-reviewed archival conference; full text verified]
     Shows that arbitrarily expressive maps can trivialize abstraction and that held-out generalization is a necessary pressure on bridge quality.
     Verbatim source quotation: “In this work, we critically examine the concept of causal abstraction by considering arbitrarily powerful alignment maps.” source

224. Sweet Rationalism. Cited resource

225. Synthbiosis. Cited resource

226. Szollosi, A., & Donkin, C. (2021). “Arrested Theory Development: The Misguided Distinction Between Exploratory and Confirmatory Research.” Perspectives on Psychological Science, 16(4), 717–724. doi:10.1177/1745691620966796

227. Thomas O. Nelson and Louis Narens. Metamemory: A theoretical framework and new findings. 1990. Primary scholarly chapter; monitoring, control, and report-method sections inspected.

228. Thompson, P., Andersson, B. J., Sundqvist, N., & Cedersund, G. (2025). “A new method for a priori practical identifiability.” PLOS ONE, 20(7), e0327593. [Peer-reviewed journal; full text verified]
     Shows how idealized measurability assumptions can inflate apparent identifiability and motivates testing what is recoverable under actual measurement constraints.
     Verbatim source quotation: “For all studied examples, traditional structural identifiability methods have widely overestimated the number of identifiable parameters.” source

229. TimeBench. Cited resource

230. Toole 2023. Cited resource

231. Vaioleti 2006. Cited resource

232. Veit, W. (2026). “Valence, Mirror Self-Recognition, and the Gradualist Evolution of Consciousness.” Philosophia, 54, 43–63. [Peer-reviewed]
     Even a defender of a stronger consciousness role distinguishes cognitive-scientific valence/value assignment from conscious realization.
     Verbatim source quotation: “Since valence, in a cognitive science framework, refers simply to value assignment, it is silent on how such valence is neurologically implemented.” source

233. Waller 2023. Cited resource

234. Walsh, D. M., & Rupik, G. (2023). “The agential perspective: Countermapping the modern synthesis.” Evolution & Development, 25, 335–352. doi:10.1111/ede.12448
235. Woodward, J. (2025). “The place of explanation in scientific inquiry: Inference to the best explanation vs inference to the only explanation.” European Journal for Philosophy of Science, 15, 21. [Peer-reviewed]
     Strong contemporary challenge to unrestricted inference-to-the-best-explanation; retained as a live rival rather than treated as consensus.
     Verbatim source quotation: “It argues that IBE is not a defensible form of inference.” source
236. Ylikoski, P., & Kuorikoski, J. (2010). “Dissecting explanatory power.” Philosophical Studies, 148, 201–219. doi:10.1007/s11098-008-9324-z

237. Yubin Kim and colleagues. Capable language models can outgrow the benefits of collaboration. 2026, published 24 July. Peer-reviewed controlled study; primary methods and limitations inspected and publisher page rechecked through its cookie-error URL.

238. Şerban, M. (2026). “Rethinking interlevel experiments: no remainder from evidence for causal relations.” European Journal for Philosophy of Science, 16, 12. [Peer-reviewed]
     Counterpressure in the opposite direction: matched interlevel evidence may support causal mediation without a distinct constitutive remainder.
     Verbatim source quotation: “Drawing on paradigmatic action potential experiments, I demonstrate that practices satisfying MIE’s formal requirements consistently establish causal mediation relationships without requiring constitutive interpretation.” source