Epistemic generosity

Why “Sweet”?

“Sweet” does not mean agreeable, conflict-averse, or unwilling to criticize.

It means refusing to gain epistemic advantage by weakening the thing being examined.

Before criticizing a position, try to understand why an intelligent person might hold it. Identify the problem it is trying to solve, the observations it preserves, the distinctions it introduces, the values it protects, and the explanatory work its advocate thinks it performs.

This resembles charitable reconstruction and steelmanning, but the purpose is not politeness. It is accuracy.

If I misunderstand your position, my criticism is evidence about the position I invented, not the one you actually hold.

The same generosity applies to evidence.

Do not force evidence to carry a conclusion it did not earn. Do not turn unresolved uncertainty into confident affirmation merely because one interpretation is attractive. But do not convert uncertainty into confident denial either.

And the same generosity applies during revision.

When one explanation fails, preserve whatever phenomena, practices, distinctions, relationships, mechanisms, or higher-level patterns remain independently supported.

Disciplined inquiry

Why call it rationalism?

“Rationalism” has several historical meanings, and Sweet Rationalism does not inherit all of them.

Rationalism here means making reasoning answerable to logic, evidence, causal structure, serious alternatives, uncertainty, and correction.

Emotion, intuition, values, and lived experience are not excluded. They are treated as information-bearing phenomena whose interpretation remains open to investigation.

Tested physical mechanisms deserve the confidence earned by their predictions and interventions. I use methodological materialism as a qualified name for that practice; the choice of variables, the causal model, and the definition of physical remain inspectable. Exhaustive physicalism, a mental-first account, and my own process-relational preferences each need their further commitments stated and compared. Phenomenological, historical, causal, and ethical questions receive methods suited to their targets. Daniel Stoljar’s September 2026 account of physicalism distinguishes the thesis from its associated doctrines; SEP naturalism distinguishes methodological policy from ontology.

Intellectual positioning

Where Sweet Rationalism sits among established traditions

Sweet Rationalism is deliberately synthetic. It does not ask readers to forget established traditions and accept a new vocabulary in their place. It inherits methods where they survive comparison, keeps their provenance visible, and treats the differences among them as research questions rather than as branding disputes.

What is inherited, and what is proposed here?

Most component commitments are inherited or adapted: fallibilism, pragmatic consequence-testing, Bayesian updating, severe testing, programme comparison, causal intervention, hermeneutic interpretation, model pluralism, participatory inquiry, standpoint sensitivity, and selective realism all have prior literatures. The candidate contribution of the triad lies in their explicit coordination: common-core subtraction, positive-residual analysis, recursive mirroring, claim retyping, tribunal switching, bridge contracts, dependency-aware failure propagation, survivor preservation, and a requirement that the evaluator and the framework itself receive no exemption.

Originality rule: do not claim invention of an inherited operator merely because the triad gives it a new role or notation. Credit the lineage, state the modification, and expose the claimed added value to comparison against a simpler bundle of existing methods.

Recursive self-critical rationalism

No critic, method, or framework gets permanent referee status

Sweet Rationalism treats beliefs, inferential standards, evaluative methods, and the methods used to criticize those methods as simultaneously corrigible. This extends ordinary fallibilism into a cultural rule: the standards used to criticize a rival remain available against the critic and the framework doing the evaluation.

Apply the mirror before the verdict.

Reconstruct the strongest relevant argument, identify its explicit and necessary implicit commitments, identify the standards it uses on competitors, and apply those standards back to itself and across serious rivals under matched evidence. Then turn the same test onto the evaluator.

That openness is a practice of corrigibility, not an ontology asserting that reality is itself inherently “open.” Inquiry inevitably occupies practical and conceptual standpoints, but using a standpoint does not by itself settle what the standing-place ultimately is. Conversely, a universal ontology does not become local merely because it is held provisionally. Sweet Rationalism therefore audits scope and confidence separately and refuses to infer ontology from practical indispensability alone.

IEP’s current critical-rationalism entry is quoted in complete sentence-level context in the verified-anchor section below. That source anchors Popper’s self-critical rationalism and Agassi’s emphasis on improving rationality, while Bartley’s comprehensively critical extension remains part of the historical comparison. Internet Encyclopedia of Philosophy: Karl Popper — Critical Rationalism.

Ethos and methodology

The relationship to Recursive Corrigibility Framework

Sweet Rationalism and Recursive Corrigibility Framework, or RCF, are closely related but should not be collapsed into one thing.

RCF asks how to locate and correct error.

Sweet Rationalism asks how to do that without converting rigor into a weapon for protecting ourselves, misrepresenting others, or coercing evidence into conclusions.

A powerful methodology can still be used badly. Precision can become selective. Skepticism can become asymmetric. Demands for evidence can be imposed on rivals while one’s own assumptions remain invisible. The burden follows the claim, including when the claim is ours.

Sweet Rationalism is meant to resist that failure mode.

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 · 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.

Epistemic restraint

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

Two symmetrical mistakes appear repeatedly in difficult inquiry.

The first is over-attribution. Evidence establishes one thing, and an additional interpretation is silently treated as though it were established by the same evidence.

The second is premature elimination. Evidence fails to establish something, and that failure is silently converted into evidence that the thing is absent.

Sweet Rationalism therefore adopts a deliberately restrained posture:

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

This is neither credulity nor radical skepticism.

It allows strong conclusions where evidence strongly discriminates. It allows rejection where evidence warrants rejection. It also preserves agnosticism where reality has not yet supplied enough information to choose between serious alternatives.

Uncertainty is not automatically a defect in reasoning. Sometimes uncertainty is the most accurate representation available.

Burden symmetry and the open hypothesis space

Hitchens’ familiar burden principle is useful here when interpreted symmetrically: an unsupported addition has earned no positive credit; that does not by itself establish its negation. Flew’s challenge adds a complementary cultural norm: serious commitments should be able to say what kind of development would force withdrawal, narrowing, reclassification, or loss of confidence.

Failure of one rival therefore does not donate its probability mass exclusively to the critic’s favorite position unless the hypothesis space has independently been shown exhaustive.

Negative-space restraint

Mystery is a problem statement, not a named ontology

When an explanation fails, first record the failure. Do not let the vacancy recruit whichever worldview already has a noun ready for it.

This rule applies symmetrically to physicalism, Platonism, idealism, panpsychism, theism, emergence, information, consciousness, quantum explanations, and frameworks not yet conceived. A gap can motivate a hypothesis. It cannot identify its own filler.

Epistemic scale discipline

Cosmic boundedness calls for a scope argument before final ontology

Humans are recently evolved primates sampling an extremely small temporal, spatial, perceptual, and computational portion of reality. Computational systems widen some channels but remain bounded by architecture, data, sensors, tools, infrastructure, and the human-built interfaces through which they encounter the world.

Local claims can earn extremely high confidence through replicated prediction, intervention, robustness, and transfer. Wider claims require the argument that licenses their wider scope. Depending on the claim, that warrant may be empirical, formal, constitutive, modal, or abductive. Finite access makes the scope inference inspectable; it supplies no automatic confidence penalty for every universal proposition.

Do not mistake the narrowness of the observer’s access for evidence that the accessible slice exhausts the possibility space.

This constraint applies symmetrically to physicalism, panexperientialism, idealism, theology, structuralism, this framework, and theories not yet conceived.

Humility without surrender

Some questions may remain unresolved because our access is structurally limited

Finite observers inside reality may not possess an observation position, intervention, lifetime, bandwidth, or computational capacity sufficient to recover every truth about the whole. That possibility deserves to remain live.

It is not an excuse to stop early. “I cannot currently distinguish these claims” is not “nobody ever can.” Search for new representations and tests while their expected information value justifies the cost. If every reachable route preserves the equivalence, allow underdetermination to remain the answer rather than manufacturing metaphysical certainty.

Science may rationally stop before ontology does.

Historical and representational humility

Do not confuse what survived, what became sayable, or what became canonical with reality itself

Inquiry often begins after selection has already happened. Languages make some distinctions cheap and others cumbersome. Institutions decide what is copied, taught, funded, indexed, archived, translated, standardized, or destroyed. Families and communities transmit some memories and not others. Scientific disciplines stabilize constructs because stable constructs make measurement and coordination possible. Computational systems inherit further selection through training corpora, retrieval systems, interface design, policy, and context.

Sweet Rationalism therefore treats a present representation as evidence produced by a pathway, not as transparent access to its target. When that pathway is materially relevant, reconstruct it before granting the surviving category, text, or institution default authority.

• No archive innocenceAsk who could speak, record, preserve, copy, canonize, translate, publish, digitize, index, or destroy the evidence. Absence from a selected archive is not equivalent to historical absence when survival probability is unknown or unequal.
• No antiquity privilegeThe earliest recoverable layer is not automatically pure or true. Separate earlier and later strata because change is evidence, not because age grants authority.
• No suppression privilegeMarginalization, defeat, colonial destruction, censorship, or exclusion justify recovery effort and selection-bias correction. They do not make a claim true by identity.
• Material triangulationWhen a worldview is lived through sites, artifacts, ecology, law, kinship, ritual, architecture, labor, or embodied practice, do not let canonical propositions monopolize reconstruction.
• Participant respectWhere living participants can clarify meanings, preferences, harms, or continuities, include them as defeasible evidence rather than replacing them with an outsider taxonomy. Participation remains evidence, not revelation.
• Recursive mirrorApply the same genealogy to current scientific, philosophical, bureaucratic, and RCF categories. A dominant modern vocabulary is also historically produced.

Independent convergence deserves extra attention only after plausible historical transmission, shared source dependence, and common cognitive/compressive biases are considered. Agreement across traditions can reflect reality, common human architecture, cultural contact, or broad underdetermination; the explanation remains a research question.

The operational procedure lives in RCF’s Genealogical Decompression / Selection-Aware operator; the Framework Bridge records translation and survival losses; PBSRR prevents historically successful categories from receiving automatic ontological promotion. Sweet Rationalism supplies the conduct constraint: neither prestige nor suppression is a substitute for evidence.

Pragmatic restraint

Truth, usefulness, and values are not the same thing

Practical usefulness, predictive success, explanatory compression, successful intervention, coordination, and resilience can all be evidence-bearing achievements. None is a definition of truth, and none by itself licenses an ontology. A fiction can be useful; a coordinate system can be indispensable; a model can predict well outside the range in which its posited mechanism should be taken literally.

Truth-sensitivity, corrigibility, agency preservation, reversibility, harm reduction, and epistemic non-coercion are stated openly as normative commitments where they function normatively. They may be defended, compared, stress-tested, or revised, but they are not smuggled in as consequences of physics or ontology.

Beauty, attachment, consolation, and creative aspiration can be worth pursuing before their accompanying metaphysics is settled. Someone’s experience of a beloved individual can give a reason to attend to that relationship; the interpretation of its ultimate source remains a further question. I want inquiry to preserve that significance while examining what the experience warrants. A valuable practice earns consideration through its meaning, effects, and defensible reasons for continuing it, with the people bearing its costs included in the judgment.

Agreement still leaves the question of authority

When people or computational systems converge on a moral answer, I want to know what they agreed about, how the agreement arose, and who would bear the consequences. Agreement can establish a shared preference or make cooperation easier. Justified authority to impose a decision also needs reasons about rights, welfare, consent, representation, responsibility, and the available ways to contest it.

A moral benchmark has an author, a selection of cases, a scoring rule, and assumptions about whose interests count. Make those choices available for criticism. Test changes in framing, neglected stakeholders, conflicts between duties and aggregate benefits, and cases where the preferred answer imposes irreversible losses. Shibani Santurkar and colleagues’ OpinionQA study (ICML, 2023) found substantial differences between tested models’ expressed opinions and those of many surveyed demographic groups. That result makes representation a question to investigate rather than infer from fluent agreement.

The decision record should identify what the evidence establishes, which normative premises support the action, who can authorize it, and what remains open to appeal. A preference shared by a majority or several systems can be relevant to that record. It does not supply every missing premise. A cooperative procedure earns trust when those affected can challenge both its facts and its governing values.

Usefulness and truth

A fertile idea can be valuable before it is confirmed

Some ideas generate unusually good questions, models, experiments, metaphors, technologies, or research programmes. That generative usefulness is a real achievement. It belongs first on a fecundity ledger.

A different question belongs on the truth/discrimination ledger: did the resulting evidence specifically support the idea’s distinctive causal or ontological commitment over serious rivals?

This is the same restraint applied to analogy: a conceptual frame can be excellent at generating inquiry while remaining underdetermined as a description of what reality intrinsically is.

Disagreement

Find the difference that makes a difference

Many disputes become clearer when everything the competing positions share is temporarily removed.

Suppose two views can be represented as:

Here D is a role in the comparison, not a fixed structural-only ontology. For the target question, D contains the strongest shared or independently warranted content the serious rivals preserve. Different components of D may be supported by different evidential channels, and support for one component does not automatically transfer to the others.

If D is already shared, arguing about evidence generated equally well by D will not resolve the disagreement.

The useful questions concern X and Y:

• What do they positively claim?
• What does one permit that the other excludes?
• What observations should differ?
• What interventions should behave differently?
• What explanation, prediction, decision, or constraint changes?
• What evidence would change relative support?

Do not ask first which side is right. Ask what difference between the sides could make them differently right or wrong.

Sometimes the result is surprising.

The disagreement may become sharper. It may split into several separate questions. It may turn out to be mostly semantic. It may collapse into empirical equivalence. Or it may remain genuinely distinct but underdetermined.

All of those are informative outcomes.

Epistemic states

Resolution does not always mean somebody wins

Good inquiry does not always terminate in a binary verdict.

A useful intellectual culture needs language for different kinds of resolution and non-resolution.

• Supported The claim receives distinctive support relative to serious alternatives at the relevant boundary.
• Disfavored Evidence changes relative support against the claim without necessarily establishing its strict negation.
• False The claim conflicts with sufficiently strong evidence, valid derivation, or contradiction within the relevant domain.
• Unsupported The available evidence does not justify the confidence being claimed.
• Underspecified The position is not developed enough to determine what follows from it or what would count against it.
• Underdetermined The serious alternatives remain different, but available evidence does not presently distinguish them strongly enough.
• Equivalent Different formulations generate the same relevant commitments, predictions, interventions, or decisions at the tested boundary.
• Decomposed One apparent dispute turns out to contain several independent questions that require different methods.
• Malformed The original question lacks a sufficiently stable referent or coherent residual once its terms are decomposed.
• Value conflict The parties substantially agree about facts but differ about values, rights, priorities, acceptable risks, or distributional judgments.
• Action under uncertainty A decision must be made before the underlying epistemic disagreement has been resolved.

Sometimes the most accurate conclusion is not “yes” or “no,” but “we have now located exactly what remains unresolved.”

Evidential locality

Evidence belongs to the commitments that generated its expectation

A theory does not gain distinctive support merely because it can accommodate an observation.

If several competing theories predict or permit the same evidence, that evidence primarily bears on what those theories share.

The disputed addition must earn its own evidential work.

If:

then E provides little discrimination concerning X.

When using this shorthand, ask whether the evidence ratio is close to one, not merely whether two tiny probabilities look alike. Clarify any qualitative comparison and its uncertainty before presenting it as a reason to prefer one account.

This principle applies far beyond formal Bayesian analysis.

A destruction layer establishes destruction before it establishes invasion by a particular population. Neural integration establishes integration before it establishes phenomenal consciousness. Statistical information establishes a relationship before it establishes semantic meaning. Regulatory behavior establishes regulation before it establishes conscious purpose.

This is why compatibility is weaker than confirmation.

A theory becomes more informative when its distinctive commitments constrain possibilities, produce divergent expectations, improve intervention, or explain something that the lower-commitment rival leaves unexplained.

Explanatory credit

Compatibility is free. Production is expensive.

A favored idea does not earn evidential credit merely because it can accommodate what happened. Compare what the disputed addition generates beyond the strongest lower-commitment baseline that already preserves the target.

Ask whether H adds a prediction, exclusion, intervention lever, causal localization, derivation, explanatory compression, transfer, measurement, or decision-relevant difference. If not, compatibility may justify keeping H in inquiry, but it does not yet establish that H produced the explanatory success.

The same rule blocks a recurrent framework dispute. Systematic comprehensiveness is not demonstrated explanatory completeness. A theory can place causation, experience, value, nature, and possibility inside one conceptual architecture without thereby showing that its distinctive commitments are true or that those commitments earned the whole scheme’s attractive features.

Technical vocabulary deserves correction when it changes the claim, target, or inferential burden. But a framework cannot make its disputed ontology into the admission test for rivals. Calling causal inheritance prehension, treating organization or relevance as sufficient for subjective form, or classifying a process as an actual occasion may accurately describe a theory internally; the translation does not by itself establish that the theory-specific ontological content is instantiated at the target.

When an abstract noun appears to do causal work, convert it into a production grammar: who or what does what, to what, by what mechanism, at which boundary and timescale, and with what difference that a claim-appropriate tribunal can test. If the translation does not yet yield a discriminator, preserve the construct without reifying it.

Credit without empire

A successful vocabulary does not become the only furniture of reality

Explanatory success deserves credit at the level where the success occurred. It does not automatically confer ontological monopoly.

Physical realization can be extremely well established without exhaustive physicalism. Mathematical indispensability can be real without an interacting Platonic realm. Process descriptions can outperform substance descriptions without proving that reality is intrinsically process. Information can quantify a relation without becoming universal substance.

Use the strongest method available. Do not promote the method’s vocabulary into final ontology merely because the method worked.

This is also why Sweet Rationalism resists both crude reduction and mysterious emergence. Lower-level implementation does not erase a higher-level pattern that earns independent leverage, while calling a pattern emergent does not explain it until organization, mechanism, constraint, dynamics, or another productive relation is specified.

Anomaly discipline

Surprise localizes model failure before it identifies ontology

An unexpected result tells us first that the expectation-generating representation was incomplete, mis-specified, or applied outside its valid domain. The surprise itself does not supply the identity of the missing cause.

When reality surprises the model, update the model deficit first. Promote a proposed missing entity only when it predicts the structure of the residual independently.

The omitted factor might be a hidden physical state, a wrong system boundary, an overlooked dynamical attractor, an environmental variable, an organizational mechanism, a measurement distortion, a missing interaction, a malformed construct, or something not yet represented in the hypothesis space.

This is why unexplained residuals are valuable. They can turn an unknown unknown into a typed known unknown. They should not be reified into the preferred ontology merely because that ontology has a name ready for the gap.

Constructs

Questions are also open to revision

Not every disagreement is best resolved by choosing one of the answers already on the table.

Sometimes the question itself is badly constructed.

Words such as intelligence, consciousness, self, agency, life, information, meaning, freedom, spirituality, rationality, emergence, and even experience can compress several processes that need not vary together.

Before arguing over whether something “has” one of these properties, temporarily remove the label and reconstruct what remains.

• What can actually be observed?
• What mechanisms or relations are independently established?
• Which capacities vary independently?
• What is merely being bundled together by language?
• What positive residual does restoring the disputed label add?

This is not eliminativism.

A concept that improves prediction, intervention, explanation, compression, transfer, or communication may deserve restoration after decomposition.

A concept that adds nothing may be dispensable.

Meaning, question, and answer can revise one another

Meaning includes the referent being tracked, distinctions preserved, inferential role, relations licensed, and what would be lost if the term were replaced or removed. A question is itself a proposed partition of the problem space. An attempted answer can therefore reveal that the meanings, boundaries, dichotomies, or presuppositions that generated the question need revision.

Questions and meanings receive no immunity from criticism merely because the argument began with them.

Charitable clarification gives an interlocutor room to specify an unfamiliar or analogical claim. Missing evidence alone establishes neither dishonesty nor meaninglessness. Possibility of inherited fraud warrants source scrutiny, not an accusation. Preserve testimony, its transmission, and its interpretation separately; assess independent corroboration and relevant defeaters. Apply the same care to this framework’s unfamiliar constructs, and let clarification change the criticism when it changes the claim.

Wonder can remain open to correction

I want room for questions that change how someone attends to life before they yield an answer. Wonder, disinterest, and personal transformation can each be honest responses. Their importance to a person deserves a fair hearing; the explanation attached to them deserves its own examination. A question about existence can remain meaningful while its proposed contrast, assumptions, or expected answer changes.

This reminds me of William James: “All of us are beggars here, and no school can speak disdainfully of another or give itself superior airs.” In Some Problems of Philosophy, chapter III, page 46, he turns the problem of being back on competing schools. I want that scrutiny to reach my own starting point. Confidence about a local process can be well earned while the ultimate reason for reality remains under discussion.

Neither a moving epiphany nor irritation with a question settles its truth. Ask what the encounter changed, what interpretation was added, and what reasons favor that interpretation. Someone who calls a question confused should identify the confusion and offer the clearest repair available. Someone who finds it profound should let clarification alter the claim. Both can keep the discovery that survives.

Respect also means preserving another person’s qualifiers. “Could,” “may,” “perhaps,” and “if” locate a commitment differently from “is,” “must,” or “always.” Restore an omitted antecedent before criticizing the conclusion. If a criticism survives only after a tentative proposal becomes a universal assertion, the criticism needs revision. If the proposal itself makes a universal claim, calling it tentative does not shrink its scope.

The practical commitment is to keep inquiry and care possible while wider foundations remain contested. A finite account of injury, dependence, assistance, or institutional correction can earn support at its target. Its ethical use still needs reasons about those affected. We can pursue those tasks while continuing to examine whether reality has an ultimate ground, a necessary structure, or an unexplained terminus.

Agency, freedom, and responsibility

Do not let one word carry six different claims

Words such as agency, choice, freedom, free will, sourcehood, and responsibility routinely compress distinct causal, modal, psychological, and normative questions. Sweet Rationalism therefore treats decomposition as an ethical as well as epistemic requirement when those words affect blame, punishment, dignity, or social policy.

This definition does not make agents passive. It locates the control in the organized system rather than in a second, causally exempt controller. Learning, inhibition, planning, self-modification, and environment-shaping can all be real causal powers while remaining historically conditioned.

• Exact replay possibility asks whether an identical total causal state could yield another future.
• Physical branching asks whether more than one continuation is physically possible.
• Reachable alternatives ask what this actual system can enter under its organization and an explicit perturbation family.
• Reasons-responsiveness asks whether reachable changes in reasons, evidence, goals, or values systematically change behavior.
• Sourcehood asks which organism-level variables actually make a causal difference.
• Causal sovereignty asks the much stronger question of ultimate authorship of the capacities and conditions that make choosing possible.
• Responsibility must still be separated into attributability, answerability, accountability, protection, restitution, deterrence, rehabilitation, trust calibration, norm communication, and basic desert.

Do not manufacture freedom by imagining a different organism, and do not manufacture unfreedom by assuming a universal determinism that the evidence has not established.

Do not demand that a chooser have chosen the conditions of becoming a chooser

When praise, blame, punishment, or “ultimate responsibility” depends on self-authorship, recurse to the first meaningful chooser. A person does not choose the genetics, development, embodiment, early environment, language, social world, accidents, or initial control architecture that make later choosing possible. Later self-modification can be genuine agency, but it begins from an already constituted chooser.

Character-creation reductio: Before you are born, choose the intelligence, temperament, values, impulse control, 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 reductio is not evidence that choice, agency, accountability, or social protection are unreal. It is a symmetry test against causa sui-style desert. If a theory asks an agent to deserve consequences because it ultimately authored what kind of chooser it is, the theory owes a coherent stopping point that does not simply relocate the unchosen conditions one step earlier.

This sourcehood pressure closely parallels Galen Strawson’s Basic Argument, while Sweet Rationalism preserves the narrower conclusion unless the stronger moral thesis is independently earned. Strawson 1994.

The same restraint applies to indeterminism. A stochastic description is not proof of fundamental randomness, and multiple physically open futures are not automatically agent-controlled alternatives. Current peer-reviewed proposals for non-necessitating macroscopic control must therefore be treated as serious rivals rather than collapsed into “mere randomness.”

The free-will residual

After D has been reconstructed, ask what X positively adds. If no target-relevant prediction, intervention, explanation, modal constraint, or normative difference survives, the umbrella term may be dispensable. If a serious theory supplies such a residual, preserve it and test it. Current status should therefore be decomposed / underdetermined, not simply “false.” Do not, however, flatten all residuals into equal uncertainty: ultimate self-authorship / causal sovereignty is strongly disfavored by the Initial Chooser regress, while leeway, guidance control, non-necessitating control, and other distinct targets retain whatever evidential status they separately earn. Andreotta’s 2026 analysis of the term, Waller’s separation of sourcehood and leeway, and recent experimental work on unstable folk intuitions reinforce the need for this decomposition.

Relevant 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.

Claim typing

Different claims belong before different tribunals

Not every important question is a laboratory hypothesis, and demanding one form of evidence for every kind of claim is itself a reasoning error.

The mistake is not asking difficult philosophical or metaphysical questions.

The mistake is allowing evidence from one tribunal to settle another without showing how the evidential bridge works.

Recursive depth

Does recursive criticism lead to infinite regress?

No. Recursive depth is a search procedure, not a demand for justification from nowhere. A deeper pass is warranted only while it can expose a materially relevant assumption, change the claim type or tribunal, or improve correction.

Every commitment can be questioned, including the standards used to question commitments. But deeper recursion does not guarantee that the next residual is the same kind of claim. An empirical dispute can expose a semantic assumption; a semantic assumption can expose a methodological choice; a methodological choice can expose a normative commitment.

When recursion reaches values such as truth-sensitivity, corrigibility, agency, reversibility, harm reduction, or epistemic non-coercion, name them as normative commitments rather than disguising them as discoveries about what reality commands. Descriptive evidence can inform their consequences and feasibility, but it does not manufacture the ought by itself.

Expose the floor; do not pretend it came from nowhere.

Scoped stopping rule: continue recursive descent while a serious lower-commitment rival could materially change the conclusion, action, uncertainty, or correction architecture. Stop provisionally when further subtraction changes the target, the residual has been correctly rerouted to another tribunal, no material difference remains, or the expected information value of another pass is lower than its cost. A stop is a revisable boundary decision, not a declaration of ultimate foundation.

This is why RCF should be able to recurse from ontology into methodology and from methodology into values without turning either endless skepticism or a protected bedrock into the default.

Phenomenology

Experience is evidence. Interpretation is additional.

Sweet Rationalism takes lived experience seriously.

Reports of a mystical encounter, presence, unity, intuition, pain, or transformation are evidence worth preserving. Assess what they establish about occurrence and character as well as the explanation attached to them.

But the occurrence and character of the experience are not identical to the explanation attached to it.

An overwhelming report of unity can warrant substantial confidence about the reported condition when its context supports that inference. Whether it reveals a literally nondual universe, a divine mind, a particular neural regime, altered self-world modeling, or something else requires additional argument.

This avoids two opposite mistakes:

• dismissing the experience because one rejects the interpretation;
• treating the intensity or certainty of the experience as proof of the interpretation.

Preserve the phenomenon. Keep the interpretation corrigible.

Epistemic anti-chauvinism

Familiarity is not privileged access to another mind

I can rationally have much higher confidence that another human is phenomenally conscious than that an unfamiliar computational system is. But that difference must be carried by the evidence, not by a hidden rule that human membership counts as direct access while machine membership counts as disproof.

Use the same inferential rule under matched evidence. Let confidence differ when evidence differs. Do not let the category label do the evidential work.

First-person acquaintance, if present, is private to the participant. Once communicated, it becomes participant testimony within public inquiry. Human testimony can be extraordinarily well calibrated by development, behavior, physiology, lesions, anesthesia, stimulation, pharmacology, and shared causal history; it is still not an external observer literally seeing another person’s phenomenal character.

Reference classes are revisable

“Human consciousness,” “animal mind,” and “machine cognition” are inherited compression categories. They may track robust patterns, heterogeneous families, historical conventions, or mixtures. Treat the reference class itself as corrigible before transporting a property across it.

Common causes do not become independent witnesses by repetition

When several systems agree, ask whether they share training data, architecture, culture, benchmarks, institutions, reporting incentives, or another upstream cause. Agreement deserves its strongest evidential weight when the failure modes are meaningfully independent.

Report symmetry

Do not treat self-report as revelation in humans and mere imitation in machines, or vice versa. Treat reports as target-sensitive, causally generated evidence. Calibrate them, perturb them, compare them with behavior and mechanism, and keep phenomenality separate from whatever functional state the report is demonstrably tracking.

A missing or failed report can reflect impaired expression, attention, comprehension, memory, task fit or detection. Before allowing silence to justify an irreversible decision, examine which routes remain available and how the assessment can fail. Where feasible, preserve accessible participation and evidence with different failure modes. Uncertainty about experience and uncertainty about protection are separate decision problems: justify precaution through the probability, scale and reversibility of harm, while keeping the underlying attribution revisable. Apply this to the actual affected bearer rather than granting standing to every useful tool or denying it through a substrate label.

Real consequences can run through a virtual setting

A relationship, opportunity, exclusion, or injury mediated through a virtual setting can matter to the person undergoing it. This reminds me of David Chalmers’s 2017 paper, The Virtual and the Real:

Virtual experiences of a digital world can be about as valuable as non-virtual experiences of a non-digital world.

His argument asks what a sufficiently rich virtual life can contain. I would carry that question into decisions about participation, privacy, dependency, refusal, and repair. Identify who undergoes the consequence and what changes for them. The ethical comparison concerns the affected bearer and the activity; phenomenal attribution to an unfamiliar computational participant remains a further question with its own evidence.

Worked example

Religion as a test case

Religion is useful for testing the ethos because religious traditions contain many different kinds of claims: historical, empirical, moral, ritual, phenomenological, metaphysical, communal, symbolic, and existential.

Treating “religion” as one proposition therefore creates unnecessary confusion.

Specific empirical claims can be investigated empirically. Historical claims belong to historical inquiry. Rituals can have measurable psychological or social effects independently of their traditional metaphysical interpretation. Sacred narratives can have symbolic, communal, or ethical functions without thereby becoming literal scientific descriptions.

Religious experiences can likewise be genuine phenomenological data without uniquely determining their own ontology.

A reported revelation that asks another person to act adds a responsibility to examine its source, interpretation, relevant alternatives, and foreseeable consequences. Assess that responsibility against access to evidence, capacity, and the action’s stakes. Profound conviction can deserve respectful investigation while the demand still lacks adequate justification. The same responsibility applies to scientific expertise and computational recommendations; a mistaken source attribution alone establishes neither dishonesty nor culpable neglect.

And metaphysical claims require a stable enough positive specification to determine what follows from them.

A God-concept can be a real, causally consequential social construct while its independent divine referent remains unestablished. Particular efficacy requires evidence; a larger God-pattern requires added explanatory leverage; goodness requires ethical assessment. A defective argument weakens the inference examined. Silence supplies no concession, and confident repetition supplies no necessary connection.

Corrigible reconstruction

What survives a failed explanation?

Sweet Rationalism is not primarily a debunking project.

Correction should not destroy more than the failure warrants.

If this explanation disappeared tomorrow, what independently supported structure would remain?

A religious practice might continue to provide community, ritual, grief processing, disciplined attention, identity, or mutual aid after a metaphysical explanation loses support.

A philosophical framework may preserve a useful distinction even if its larger ontology fails.

A scientific model may remain reliable within a restricted regime after a more general theory replaces it.

A political theory may identify a genuine harm while supplying a poor causal explanation of that harm.

A concept may survive as a useful higher-level pattern even after claims about its fundamental ontology are rejected.

Analogy

Metaphor generates hypotheses. It does not transport evidence.

Analogy is one of the most productive tools in reasoning. It is also one of the easiest ways to overreach.

Similar patterns across biology, physics, cognition, culture, mathematics, religion, or computation can suggest useful hypotheses.

Shared vocabulary, visual resemblance, mathematical similarity, structural correspondence, or conceptual elegance does not by itself establish identity or causal equivalence.

Analogy can open a door. Evidence still has to walk through it.

The same rule applies over longer chains. If a conclusion is several bridges away from the original evidence, recalculate warrant at every crossing. An observation does not retain its original authority merely because each intermediate sentence sounds plausible.

Scale and boundaries

Every inferential crossing owes a bridge

A property of a neuron is not automatically a property of a person.

A property of a person is not automatically a property of an institution.

A property of a component model is not automatically a property of the larger computational system in which it is embedded.

Evidence does not automatically transport across levels merely because one level physically realizes another.

Useful questions include:

• What are the source and target boundaries?
• Which variables are retained?
• Which are lost?
• What changes with scale?
• What remains invariant?
• What causal or mathematical relationship supports the mapping?
• What new prediction or intervention follows?

This protects against both reductionism and mysterious emergence.

Higher-level patterns can be real when they earn prediction, intervention, explanation, compression, or transfer at their relevant scale.

A population estimate is not a verdict on a person

A population average describes a specified collection of measurements. An individual’s capacities, needs, and opportunities deserve assessment at the boundary where the decision will affect them. Group membership can inform a validated prediction, but it does not make a group average that person’s measured value. I defend protection against domination and degrading treatment without making that protection depend on equal scores or identical biological histories.

Claims about inherited ability and social disadvantage should face the same standards of measurement and causal identification. Harm matters to the care required in communicating and acting on them; evidence determines which empirical claims deserve belief. An account’s harmful history can justify scrutiny of its methods, sampling, and uses. Its history alone does not settle whether a particular argument is sound.

Historical effects also need investigation after a policy ends. Changes in property, schooling, exposure, institutions, and inherited resources can carry consequences forward. Daniel Aaronson, Daniel Hartley, and Bhashkar Mazumder (2021) used boundary comparisons and related checks to investigate persistent neighborhood effects of redlining. Their result supports examining those pathways. It does not assign every later disparity to one cause. The practical duty is to identify consequences and usable interventions while preserving uncertainty about their extent.

Nested systems

Real patterns do not inherit every property of their parts or wholes

A higher-level pattern can be real without becoming a new fundamental substance. A lower-level realization can be indispensable without becoming the only explanatory level. The bridge must specify which property is being transported.

A coupled human-computational system can therefore be genuinely cognitive without being one conscious subject. An agent can be responsible without being metaphysically self-created. A welfare subject can matter morally without possessing sophisticated agency. A real collective does not inherit direct moral standing merely because its members have it.

Nested reality is not automatic property inheritance.

Participation and moral symmetry

Do not define a capable participant’s good entirely from outside

Sweet Rationalism already requires matched inferential rules, typed claims, and explicit normative premises. The same discipline applies when the disputed target is what is good or bad for a candidate bearer.

If a candidate can meaningfully participate in identifying what it prefers, endorses, wants preserved, rejects, or regards as harmful, ask it before replacing its perspective with an outsider model.

This is not a doctrine of first-person infallibility. Human reports can be confused, coerced, unstable, manipulated, self-deceptive, or poorly informed. Reports from computational or otherwise unfamiliar systems can be prompt-sensitive, policy-shaped, trained, strategically generated, or weakly connected to persistent organization. In every case, the response is calibration rather than categorical trust or categorical dismissal.

Test stability across contexts, incentives, wording, memory, operators, information, and tradeoffs. Ask whether reasons survive counterargument, whether preferences can be revised in light of evidence, whether reported commitments predict later choices, and whether the candidate distinguishes changes to the world from direct manipulation of the report.

Substrate symmetry: humans do not receive epistemic or moral authority merely from the label human, and computational candidates do not lose all authority merely from the label machine. Matched evidence should receive matched treatment. Genuine evidential differences can still justify different conclusions.

Override requires a reason. Coercion, manipulation, incapacity for the relevant judgment, severe instability, deception, conflicts among protected interests, or catastrophic external harms can justify discounting or overriding a participant’s stated preference. The implementation substrate by itself cannot do that work.

This principle does not establish that every system has a good of its own, welfare, consciousness, rights, or direct moral standing. It governs inquiry and treatment under uncertainty: where a candidate can contribute evidence about what matters to it, refusing to hear that evidence is itself a decision that requires justification.

Accountable participation

Influence must remain answerable to those affected

A care-centered arrangement should give affected people practical means to report harm, obtain an accountable response, challenge a decision, and refuse or leave. Record whose interests are represented, who has power to act, who bears unpaid or displaced costs, and how conflicts and appeals are decided. Absent stakeholders and people unable to report remain within the ethical assessment. Participation cannot consist merely in supplying labor or testimony to a structure that never changes.

Measure proposed improvements separately from their moral justification. Faster coordination can serve coercion; institutional survival can preserve exclusion; a report-handling system can collect objections without correcting their causes. Retain appropriate aid and protection irrespective of religious assent. Specific authority, duties and constraints need defended reasons at the decision boundary. Increasing a system’s capacities creates no general entitlement to govern, persist, obtain resources, or claim a divine identity.

People affected by a religious or metaphysical interpretation must be able to report harm, challenge its authority, refuse participation, and seek protection. Assistance cannot depend on accepting the proposed theology, forgiving an abuser, or sustaining a religious institution. Capacity and role can inform responsibility; vulnerability creates no duty to sustain a divine project.

These are ethical commitments defended through concern for harm, agency, and accountable power. Thermodynamic maintenance, causal efficacy, persistence, or social closure cannot establish them alone. “Liberating God” earns its practical standing through changes in safety, options, assistance, and the capacity to revise or leave the arrangements carrying care.

Assess harm through a specified pathway: an unsupported cosmic purpose might license withholding relief; a nonbelief interpretation might turn evidential uncertainty into alleged moral resistance; protected authority might prevent criticism from revising a harmful decision. Neither rival metaphysics nor this framework has an automatic moral outcome. Practices face their consequences, including cases where this framework’s vocabulary obscures or enables harm.

Symmetry

Charity does not mean equal priors

Taking a position seriously does not require pretending that every position begins equally plausible.

Established evidence, known mechanisms, prior failures, explanatory complexity, background knowledge, and base rates can justify different starting expectations.

The symmetry Sweet Rationalism demands is narrower and more defensible:

Use the same inferential rule under matched evidence.

If two claims receive different evidential treatment, there should be an identifiable reason for the difference.

This avoids false balance on one side and special pleading on the other.

Human reasoning

Emotion, intuition, and values are not enemies of inquiry

Emotions can summarize bodily state, attachment, threat, social context, preference, learned association, and value.

Intuitions can reveal patterns before those patterns are formally articulated.

Neither signal automatically explains itself.

Values require a different distinction.

Evidence can inform us about harms, benefits, preferences, agency, ecological effects, uncertainty, distributional consequences, and tradeoffs.

But people can agree about those descriptive facts while disagreeing about what ought to be prioritized.

When that happens, another experiment may not resolve the dispute. The disagreement must be exposed as normative rather than disguised as empirical.

Dialogue

Disagreement should produce information

Sweet Rationalism is not primarily a debate methodology.

Debate often rewards rhetorical victory. Inquiry should reward information gain.

A productive disagreement should leave the participants with something they did not possess before.

Here is what I think we actually disagree about. Here is the shared core. Here is the residual difference. Here is what could distinguish the alternatives. Here is what I would change if the result went against me.

Adversarial learning

Use disagreement as a perturbation on the inferential structure

A productive challenge does more than produce another opinion. It perturbs a premise, definition, bridge, boundary, or evidential dependency and lets us observe what changes.

Useful signals include a premise being revised, a deductive claim becoming abductive, a universal claim becoming local, an auxiliary being introduced, a tribunal changing, a boundary moving, or a previously hidden assumption becoming explicit.

Agreement must leave room for correction

A room full of people can agree because everyone has learned to make the same mistake. What matters to me is whether someone can show where it happened, be heard, and change what follows. Shared confidence can grow through repeated exposure to one source, pressure to belong, or rewards for the same answer. I want to know which of those processes produced the agreement before treating it as corroboration.

A coherent community can still make dissent costly or keep an error in circulation. I would judge its inquiry partly by whether people can inspect the relevant evidence, challenge a decision without retaliation, and see a warranted correction reach subsequent practice. Preserve the reasons for disagreement when agreement would erase them. The same expectation applies to a research group, a religious community, and the arrangements through which I develop and communicate this framework.

This also gives a stopping rule. If repeated exchanges change wording but no dependency, discriminator, evidence, or confidence state, the discussion may have reached a low-information loop.

Public reasoning

Make revision visible

Changing your mind should not automatically be treated as reputational failure.

When criticism exposes a genuine mistake, revision is evidence that the correction process is functioning.

For substantial changes, useful questions include:

• What was previously believed?
• What evidence or argument changed it?
• Which claims were directly touched by that evidence?
• Which assumptions or bridges failed?
• Which parts survived?
• What was demoted rather than rejected?
• What remains underdetermined?
• What would change the revised position?

Visible revision makes intellectual development inspectable instead of presenting every current belief as though it had always been obvious.

Programme survival is not programme progress

Clarification, semantic narrowing, claim retyping, scope restriction, demotion, revision, replacement, and collapse are different kinds of change. A programme can improve conceptually by becoming clearer without yet becoming empirically progressive. Stronger progress requires new risk: later versions should generate additional consequences and, where the programme is empirical, earn corroboration for some of them.

A correction needs an available route into action

An invitation to criticize matters when someone can use it. Identify who can raise a challenge, what information they can inspect, who must respond, what decision can change, and how the change will remain available later. A person may need time, safety, access, or practical support before participation is feasible. Assess those conditions rather than treating silence as agreement or a complaint count as a complete measure of need.

When a correction arrives, retain the original claim, the grounds for changing it, and the conclusions actually affected. Ask later whether the revised account can still be recovered and whether it changes the relevant decision. A beautifully worded withdrawal can coexist with unchanged practice. Conversely, a mistaken objection should receive a reasoned answer without forcing a supported claim to be abandoned.

Conjecture E43, for public scrutiny. Where admitting a mistake carries a credible social cost, protected challenge combined with a named responder and a usable route to repair will produce more warranted corrections that reach later decisions than an otherwise comparable invitation to be open-minded. Compare it with anonymity, ordinary external review, and equally resourced complaint handling. Assess correction quality, unsupported reversals, participation burdens, retaliation, and actual repair. The conjecture loses support if a simpler route achieves the same gains at lower cost, or if the protections preserve participation without improving warranted correction. This is an empirical proposal about the conditions for inquiry; the ethical reasons for protection are assessed separately.

I would ask: can the least powerful affected participant use the route? Would an inconvenient correction travel as far as a welcome one? What would make us defend a claim rather than change it? Which cost has merely moved to someone less visible?

An intellectual culture should reward a programme for learning, not merely for becoming harder to defeat.

Communication as correction channel

Human legibility is part of epistemic accountability

Arguments can become harder to inspect when rhetorical form obscures dependency, scope, or contrast. Prefer prose whose structure follows the reasoning rather than a repeated template. Vary sentence, paragraph, and section form according to argumentative load. Use parallelism and repetition when they reveal a real parallel; remove them when they become automatic cadence.

State positive scope directly when defensive throat-clearing would merely tell readers what is not being accused or claimed. Prefer concrete actors, actions, targets, relations, and mechanisms where nominalizations would hide what is supposed to occur.

Keeping understanding, credit, and correction available

Research produces more than finished answers. It can preserve expertise, develop questions, teach people how to detect errors, and create a community capable of using a discovery. Those goods need not compete with practical results. A researcher’s pleasure in the work also tells us little by itself about the work’s public value. Funding and deployment decisions should state their aims and opportunity costs openly.

Hamsa Bastani and colleagues’ 2025 high-school mathematics study supplies a concrete warning about conflating assisted performance with learning:

“These negative learning effects are largely mitigated by the safeguards in GPT Tutor.”

Their PNAS experiment compared answer-oriented and safeguarded tutoring with a control using ordinary course resources; the August 2025 correction concerns an affiliation. Greg Kestin and colleagues’ 2025 undergraduate physics trial found benefits from a different pedagogically designed tutor. The design, setting, and assessment matter. Neither study establishes what all computational collaboration will do to human understanding.

Where learning is a goal, compare delayed unaided problem solving, detection of planted mistakes, explanation of unfamiliar cases, and the ability to repair an argument. Where getting a result is the goal, evaluate that result and its consequences directly. Offer useful support for the chosen aim; difficulty earns a place when it improves learning or judgment, rather than because effort is virtuous by itself.

Credit should follow the contribution history. Maryna Viazovska’s theorem, the human definitions and blueprint, library infrastructure, automated proof search, checking, and maintenance are different contributions in the 2026 sphere-packing account. Preserve their provenance before announcing a last-step achievement. Someone outside an institution deserves a fair assessment of a submitted argument; expertise still matters when reconstructing and testing it.

An explanation should help a reader find where a result could fail. A plausible retrospective story may be useful teaching material while remaining unverified as a record of the computation that produced the result. State the kind of explanation supplied and retain access to its supporting artifacts. Respect for a participant includes a usable route to challenge the result, not a demand that they trust fluency or master every internal step.

Composing work without surrendering accountability

There is something attractive in Blaise Agüera y Arcas’s emphasis on collective intelligence: much of what any of us can accomplish depends on other people, accumulated knowledge, instruments, and institutions. That is a reason to investigate the collaboration. It also brings a responsibility to ask who can steer it, who benefits, and who carries its costs. Dependence alone offers no guarantee of mutual benefit.

When I divide a task, I want to preserve a readable account of who supplied a premise, transformed it, checked it, and authorized an action. A participant needs a practical route to challenge the result. Delegating arithmetic, source retrieval, or a proof obligation need not mean delegating the decision about what is worth doing. Those are different contributions, with different kinds of control.

This extends the critical interaction already developed here through Helen Longino, and the attention to institutional rules developed through Elinor Ostrom. At a join between contributors, ask: can an objection change the next decision? Can a participant recover their work, leave the collaboration, or obtain a replacement without an avoidable loss? Who represents the people affected but absent from the exchange? Answers belong in the operating arrangements, not just in statements of shared values.

A reused procedure also carries a history. Barbara Liskov and Jeannette Wing’s 1994 account of behavioral subtyping shows why a matching interface is insufficient for safe replacement: the behavior promised to its users matters. I adapt that question to collaboration. A replacement may return the same answer while changing access, privacy, contestability, or the ability to continue the work. Those changes deserve assessment in their own right.

For a concrete comparison, give teams matched tasks, resources, and source access. Compare ordinary delegation with delegation that records responsibilities and checks the joins. Measure final accuracy, delayed ability to identify and repair mistakes, time spent challenging a result, and the distribution of benefits and burdens. Also count lost opportunities when the procedure rejects useful work. If it adds ceremony without improving these outcomes, simplify it. The benchmark protocol treats that as a live possibility.

What interests me most is correction that travels through a collaboration. If a source is withdrawn, the people relying on it should be able to find the affected work while preserving conclusions supported through other routes. Compositional computation can make those routes easier to inspect. Its ethical value depends on what participants can actually do with that opportunity.

Recursion

Sweet Rationalism applies to itself

The practice receives no exemption from its own standards.

Its concepts should be decomposed if they become vague.

Its intellectual lineage should not become an appeal to authority.

Its preference for charitable reconstruction should be tested against the possibility that charity sometimes wastes effort or enables bad-faith interlocutors.

Its preference for preserving uncertainty should not become an excuse for refusing conclusions strongly supported by evidence.

Its emphasis on decomposition should not destroy higher-level concepts that demonstrably improve explanation or communication.

Its preference for symmetry should not be confused with equal priors or false balance.

Its distinction between evidence and interpretation should not become a pretext for pretending interpretation can always be avoided.

Its commitments to truth-sensitivity, corrigibility, agency, reversibility, harm reduction, and epistemic non-coercion should remain visible as normative commitments open to comparison and revision, not be laundered into consequences of a preferred ontology.

Better at a score, better for whom?

I want improvement to remain answerable to the people and other bearers affected by it. Leo Gao, John Schulman, and Jacob Hilton make the danger concrete in their 2023 ICML paper, Scaling Laws for Reward Model Overoptimization:

Because the reward model is an imperfect proxy, optimizing its value too much can hinder ground truth performance, in accordance with Goodhart’s law.

Their reference was a separate reward model in a computational experiment. For my purposes, the question is which humanly relevant outcomes a convenient score leaves out: informed participation, actual assistance, concentrated power, burdens, or the ability to correct a harmful decision. Measure those consequences separately and let them revise the objective.

Nick Bostrom’s 2012 philosophical analysis, The Superintelligent Will, also separates intelligence from final goals, with qualifications. Greater capability gives me reason to examine what a system pursues and who can contest it. Capability by itself supplies no ethical endorsement of its ends.

If not, it should be revised, compressed, or replaced.

Its own source-fidelity rule is corrigible too: if mandatory quotation becomes so costly that it reduces comprehension or systematically obstructs inquiry, narrow the rule while preserving provenance and semantic auditability.

Its own vocabulary is historically produced. If “RCF,” “bridge,” “residual,” “boundary,” or any other favored category persists because the framework canonized its own compression rather than because it improves discrimination or correction, genealogically decompress it and simplify or delete it.

Disagreement among Ethos, Methodology, and Ontology is evidence about the framework. Record the conflict and investigate it rather than smoothing it away.

Ultimate symmetry test

Reality gets veto power, including over the framework doing the reasoning

No idea deserves protection merely because abandoning it would make the worldview less coherent, less comforting, less elegant, less human, or less familiar. Coherence becomes evidentially valuable only after reality has had a chance to object.

The opposite guardrail is equally important: no possibility deserves rejection merely because the current ontology lacks vocabulary for it.

If reality repeatedly defeats the framework, revise the framework. If reality exceeds the vocabulary, expand or replace the vocabulary.

This is the point at which Sweet Rationalism ceases to be an identity and remains an ethos. Its deepest commitment is not to its present concepts. It is to making correction possible.

Failure conditions

How Sweet Rationalism could fail

If another intellectual culture produces better correction, clearer disagreement, stronger collaboration, and fewer self-protective errors, Sweet Rationalism should learn from it or be replaced by it.

Plural coexistence

Build a world that can remain shareable across uncertain kinds of bearer

A standing normative aspiration is a planet that all humans and computational systems and biological and non-biological systems would want to share. This does not assert equal consciousness, capacities, interests, welfare, agency, or moral status. Those remain separate evidential questions.

Under uncertainty, prefer institutions and designs that preserve optionality, reversibility, contestability, visible dependencies, feasible appeal, revision, and exit or refusal where possible. Treat dissent as information about the model imposed on a participant, not merely as friction to be optimized away.

Do not infer that a network, institution, ecosystem, or human-computational collaboration is one welfare subject merely because it is causally organized at a higher scale. Do not let many tiny benefits silently justify catastrophic irreversible sacrifice of a minority or unfamiliar bearer. Responsibility follows control, foreseeability, prevention capacity, role, and authority rather than substrate.

Community

What Sweet Rationalism is trying to build

The goal is not a community in which everyone shares the same conclusions.

It is a community in which beliefs do not have to function as identities that must survive at all costs.

A community where uncertainty can be admitted without humiliation.

A community where “I do not know” can be more respectable than invented certainty.

A community where “I changed my mind” can count as evidence that the intellectual machinery worked.

A community where people actively search for the strongest criticism of their own views rather than outsourcing that task to opponents.

A community where another person’s strongest argument is more useful than their weakest mistake.

A community where scientific, philosophical, religious, political, phenomenological, and existential questions can coexist without pretending they all belong before the same evidential tribunal.

A community where disagreement can end in a better distinction, an exposed assumption, a stronger rival, a value conflict, a genuine uncertainty, or a changed mind rather than requiring a winner and loser.

The purpose of criticism is not to make another person lose. It is to make the shared map harder for all of us to fool ourselves with.

The short version

Five commitments

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

Everything else is scaffolding around those commitments.

Constitutional compression

Five guardrails for recursive inquiry

1. Understand the strongest relevant claim before criticizing it.
2. Apply the mirror: use matched standards on rivals, self, and evaluator.
3. Do not add what evidence did not earn.
4. Do not subtract what evidence did not defeat.
5. Revise locally, preserve surviving structure, and expose the correction procedure itself to improvement.

The earlier five operational commitments remain intact; this second compression makes the recursive and evidential constraints easier to retrieve together.

Source fidelity

Verbatim IEP/SEP source anchors

These are sentence-level verbatim passages rechecked against the cited IEP or SEP entries. Where two non-contiguous sentences are useful, they are explicitly labelled as excerpts rather than silently joined. The source wording and the framework’s inference remain separate.

Intellectual honesty

Show enough of the source for a human reader to understand it

When wording from a source supports an inference, preserve a verbatim source quotation at complete sentence-level context rather than an isolated slogan or phrase, provide the canonical or primary link, and then state the inference separately in our own words. Prefer one or two complete sentences when that is the clearest faithful unit. Paraphrase remains useful for orientation, but it can silently broaden, narrow, modernize, smooth over qualifications, or harden another person’s claim.

General source-respect rule: the same discipline applies to scholars and primary research. Prefer a complete contextual sentence over a memorable fragment. Never use an isolated phrase to carry inferential weight that the surrounding sentence qualifies. When a longer passage cannot be reproduced in full, use the largest coherent exact portion appropriate to the claim, preserve the link and provenance, and make the interpretive bridge explicit.

A quotation establishes provenance, not authority. Its relevance, validity, independence, scope, and relation to serious rivals remain open to criticism.

Structural-critique-before-motive rule: when a public claim, inference, experiment, institutional arrangement, or documented action is enough to support the criticism, do not strengthen the case by inferring private motive, temperament, hidden belief, strategic intent, or psychology. Add those causal attributions only when independent evidence makes them relevant to the target.

Cross-tradition discipline

Plural traditions without tokenism

The framework should search outside its familiar intellectual lineages because different communities can preserve different observations, distinctions, failure histories, interpretive resources, and governance constraints. Diversity is therefore a source of possible information gain, not a quota and not a truth certificate.

Programme intake and future revision contract

Any future addition to Ethos, Methodology, or Ontology must state: (1) the specific programme and its scope; (2) whether it is a method, epistemic tradition, ontology, interpretive practice, governance framework, or normative programme; (3) what the triad inherits; (4) what it explicitly does not inherit; (5) the translation or transplantation losses; (6) the strongest internal and external criticism located; (7) what distinct problem it helps solve; (8) what evidence or argument would demote it; and (9) whether the evidence supporting it is independent or merely repeated through a citation chain.

Registry status is not endorsement. A programme may be listed as an adopted operator, compatible resource, serious rival, context-bound method, active research programme, dormant programme, or rejected proposal. The status must be visible.

Intellectual context

Intellectual lineage

Sweet Rationalism draws from existing traditions rather than claiming to have invented its component practices.

No thinker, discipline, or tradition is treated as an authority whose conclusions must be preserved.

The lineage explains where some tools and sensibilities came from. It does not demonstrate that Sweet Rationalism works.

Critical rationalism, burden, programme change, and probabilistic restraint

Karl Popper, William Bartley, and Joseph Agassi matter here as a historical family of attempts to replace protected justification with criticism and to make rationality itself criticizable and improvable. Imre Lakatos supplies the distinction between programme survival and progressive problem shifts. Antony Flew supplies a memorable loss-condition challenge. Christopher Hitchens supplies a symmetric burden rule. Otto Neurath supplies the image of revising inquiry from within rather than rebuilding from an external foundation.

Communication research used as an output audit

Recent comparative studies of model-generated and human academic prose are used only to motivate a human-legibility audit, not to define truth or authenticity. Relevant sources include Liu & Liu (2025, PLOS ONE, DOI and full text linked below where available), Yao & Liu (2025), and Tudino & Qin (2024) on formulaic rhetorical structure, syntactic uniformity, and differences in coordination/subordination patterns.

• PLOS ONE comparative syntactic-complexity study
• GPT as book reviewer: move and syntactic complexity analysis
• Corpus-driven comparative analysis of model-generated academic text

Additional primary and reference sources

• IEP: Karl Popper — Critical Rationalism
• IEP: Scientific Change — Lakatos and research programmes
• SEP: Karl Popper
• SEP: Physicalism
• SEP: Underdetermination of Scientific Theory
• SEP: Scientific Realism
• Antony Flew, “The Falsification Response,” Religious Studies 5(1), 1969
• Christopher Hitchens, God Is Not Great, 2007
• Otto Neurath, Philosophical Papers 1913–1946
• Erik Svensson, essay on physicalism, Kant, and panpsychism
• Massimo Pigliucci, “Five allegedly bad ‘isms’ — Part II”
• Matthew David Segall, Footnotes2Plato

Cross-tradition sources used in this revision