Confidence Is Not a Number You Print
A language model cannot tell a memorised fact from a plausible completion, because nothing inside it is charged with the distinction.
A language model cannot tell a memorised fact from a plausible completion, because nothing inside it is charged with the distinction. This is the architecture that would be — what it fixes, what it costs, and the two problems it does not solve.
Thesis. The failure of language models to report meaningful uncertainty is not a calibration bug to be trained out. It is the predictable consequence of an architecture in which no component is responsible for the epistemic status of a claim. A system that generates by sampling over token distributions has no place to put the answer to “how do I know this?”, and so it does not have one; the confident register of its output is a stylistic artefact of training, not a measurement. Repairing this requires a structural separation — a statistical engine that generates, and an epistemic core that governs what may be believed, what may be used as a premise, and what may be said aloud — with credence as a control variable rather than a printed decoration. I have specified that architecture across three papers. Here I set out what it delivers against the nine requirements, and where its own assumptions bind.
Abstract. I take the nine properties a trustworthy reasoning system needs — claim-level rather than token-level uncertainty; separation of epistemic from aleatoric uncertainty; source provenance; contradiction detection; reliable abstention; active information seeking; explicit tracking of assumptions; verification proportional to consequence; and calibration that survives tool use and multi-step reasoning — and show which of them have a formal answer in the architecture I have set out, and which do not. Claim-level commitment, provenance, contradiction handling, abstention, staleness decay, consequence-weighted verification and confidence propagation along inference chains all have specified mechanisms. Two do not: the epistemic/aleatoric split is absent, and the whole structure rests on the availability of an external truth signal that is thinnest in exactly the domains where the architecture matters most. I state both plainly rather than papering over them, and I identify the independence assumption in the evidence-aggregation rule as a third defect requiring repair. Two figures. Both are schematic; neither reports measurement.
I. The eight states, once more, briefly
A language model does not consistently distinguish among:
a memorised fact; a deduction from evidence supplied in the prompt; a probabilistic inference; a plausible completion; an unknown answer; a false premise embedded in the question; a conflict between sources; and an answer that was once true and may now be stale.
I set out the consequences of this in the previous essay. The short version: a competent human researcher signals which of these states they are in, in the first three words, and the signal carries the entire weight of the exchange because it tells the reader what to do next. The model produces the same prose surface for all eight, and the reader is left to reverse-engineer provenance from style — the one channel that carries no information, because style is what training optimised.
The corrective is not “reduce hallucinations”. It is that the confidence attached to a statement should predict whether the statement is true: for every credence level q the system uses, roughly q of the claims at that level should hold. Calibration.
That much is diagnosis. This essay is about the build.
II. Why this is architecture, not statistics
Start with what the generator does. A transformer defines the probability of the next token given the context window as a softmax over a learned function of that context. The output sequence is sampled from that distribution, and the sampling regime is structurally agnostic to the truth, falsity, or relevance of what it emits. There is no belief base. There is no closure under inference. There is no contradiction check. There is nothing in the operation at which the question “is this so?” could be asked, because the operation is a different question — “what follows this text?” — answered with great success.
I have called the resulting error the fallacy of mere probability: treating high statistical correlation as sufficient for belief, bypassing justification entirely. The two things are different in kind. Statistical correlation is a measure of associative regularity in a corpus. Epistemic weight is a function of the validity of the supporting evidence, the soundness of the inference that produced the claim, its coherence with the existing belief structure, and whether the system can articulate the justification at all. Statistical regularity may generate a hypothesis. It cannot constitute a belief.
The architectural response is a hard bifurcation, and it is the single most important design decision in the whole scheme:-
a statistical engine, responsible for predictive, surface-level generation — the thing that makes the text fluent, and which is genuinely excellent at that job;
-
an epistemic core, which governs belief acquisition, revision, assertion and justification tracking.
Every belief-forming operation routes through the core. No belief is formed, stored or asserted unless its provenance, support structure and epistemic status are logged and auditable. The generator may propose. Only the core may commit.
This is not a filter bolted onto a language model. A filter inspects output and blocks some of it; it inherits whatever epistemic blindness produced the output in the first place, and it cannot manufacture the missing provenance after the fact. The bifurcation instead makes the two functions structurally distinct, with different data types on either side. Fluency produces candidates. Commitment produces claims. Only the second kind has a truth value the system is answerable for.
III. From tokens to commitments
What is the unit? Not the token, and not the sentence. The unit is the proposition.
The claim object is a triple: a well-formed formula in a domain-specific language grounded in first-order logic; a temporal index marking when the claim was asserted or observed; and a contextual signature recording the ontology, experimental modality, and statistical framing under which it holds. That third element does more work than it looks. Two identical surface predicates — “X increases Y” — mean different things when one arises from a randomised controlled trial and the other from a correlational model, and a system that treats them as the same claim will aggregate what should never have been aggregated. The rule is that two claims may be treated as epistemic equivalents only if both their logical form and their contextual signature match. Semantic stratification, enforced at the representation layer rather than hoped for at the reasoning layer.
On top of that sits commitment, which is a normative state rather than a storage fact. A proposition is committed at time t when it is in the belief base, there exists a justification that entails it, and the belief base does not entail its negation. Commitment carries obligations: to accept the consequences of what one has accepted, and to revise when contradiction is derived.
The contrast with sampling is exact. Sampling maps a string to a token. Commitment maps a string to a proposition and thence to an updated belief base. One is a continuation. The other is a stance, and stances can be wrong in ways that matter, which is the entire point.
IV. Confidence as a control variable
If credence is to do work, it must license and forbid things. In the scheme I set out, propositions are tagged with a stratified confidence level, and each tier permits different operations.
Figure 1. Confidence as a control variable, not a decoration. Six tiers, from rejected to committed, and the operations each permits: only claims at Supported or above may serve as premises in standard inference; only Endorsed or above may be asserted externally; Committed claims enter deductive chains and require substantial evidential counterweight to dislodge. Everything below the premise threshold is the abstention region — and the system says so rather than generating through it. Schematic; the cut-points are conventions, the structure is the argument.
Three things follow that do not follow from printing a percentage next to a sentence.
Abstention becomes a state rather than a refusal. A claim sitting in the Equivocal band is not withheld out of caution; it is structurally ineligible to be used as a premise. The system cannot reason from it, and therefore cannot silently launder it into a conclusion, which is the usual route by which weak claims acquire unearned strength. Abstention that is enforced at the representation layer cannot be talked out of by a persuasive prompt.
Assertion rights are separated from reasoning rights. A proposition may be good enough to think with and not good enough to say. That gap — between the premise threshold and the assertion threshold — is where honest working uncertainty lives, and no system that outputs a single fluent register can represent it at all.
Transitions are governed and logged. An upward shift requires new evidence outweighing the cumulative counterweight of contradictory data. A downward shift is triggered by contradiction or by the identification of a defeater, and forces a justification audit and re-evaluation of everything downstream. A lateral conversion — from probabilistically supported to formally proven — requires verification of the whole inference chain. Credence does not drift; it moves for reasons, and the reasons are recorded.
The cut-points themselves are conventional. Parity at one half, the classical threshold at 0.95, high-confidence commitment at 0.99: these are inherited from statistical practice and are defensible rather than derived. What is not conventional is the insistence that a number in that range must correspond to a permission. A confidence score that changes nothing about what the system is allowed to do with the claim is ornamental, and ornamental confidence is precisely what we have now.
V. Provenance, and why retrieval alone does not supply it
Retrieval fixes the absence of information and introduces three failures: the wrong document; the correct document misread; and the appearance of provenance conferred on a sentence the generator composed. The third is the corrosive one, because the citation is real and survives the check the reader knows how to perform.
Provenance done properly is not a footnote. In the architecture, every claim carries a canonical signature derived from its normalised logical form, the canonical identifier of the asserting author, and a timestamp — hashed, so that logically identical claims collide to the same signature and semantically distinct ones do not. Authors carry canonical identifiers built from persistent identity records and disambiguated across naming variants, so that a claim’s lineage cannot be broken by the ordinary bibliographic chaos of initials, transliterations, and affiliation changes.
Each ingested document carries a metadata record: the canonical author identifier, the canonical claim signature, a persistent digital object identifier, a timestamp, a venue vector, funding disclosure, replication status, and a structural completeness flag. A record missing any of the essential referential anchors is not admitted. Every record must verify against a cryptographic hash and a signature from a trusted identity provider, and where the hash is committed to a publicly verifiable append-only ledger, the provenance becomes tamper-evident: any alteration to the content changes the hash and breaks the signed record.
That last mechanism is the one people find excessive until they consider what it buys. Not security theatre — auditability of the reasoning itself. If every belief traces to an immutable origin, then post-hoc rationalisation becomes structurally detectable. A system cannot quietly reconstruct why it believed something; the record of why it believed it was written when it believed it, and cannot be revised to fit the conclusion. Epistemic drift and retrospective justification are the two ways reasoning systems — human institutions very much included — decay while appearing sound. An append-only justification chain closes both.
The consequence for the reader is simple. Provenance means: this specific proposition, supported by this specific passage, from this specific verified record, at this specific time. Anything weaker consumes suspicion without earning it.
VI. Contradiction is not noise
In a system that reasons deductively, a contradiction is not a data-quality problem. From a proposition and its negation, anything follows. A belief base containing both is not slightly degraded; it is logically trivial, and every subsequent inference drawn from it is void. Contradiction is therefore not a bug to suppress but a proof that the internal model is broken and must be repaired or abandoned.
Detection is a predicate over the belief base, and resolution is an operator that restores consistency with minimal loss of information — contraction to remove the minimal subset entailing the offending claim, then revision to install its replacement. Which of the two contradicting claims goes? Not the one that arrived later, and not the one with the lower raw probability. Each proposition carries a provenance tag recording source, time and confidence, and a dominance ordering over provenance determines the outcome. Prioritised belief revision, on the strength of the record rather than the order of arrival.
At network scale the same logic goes further. Claims sit in a graph; contradiction edges connect incompatible pairs; and a cluster’s instability score measures the density and severity of contradiction among its highly believed members — the sum of the products of the credences across contradiction edges, normalised by cluster size. Above a threshold, the cluster is partitioned by a minimum cut that minimises contradiction flow across the boundary, and the offending region is placed in epistemic quarantine: the minimal subset whose isolation restores coherence. Propagation into and out of a quarantined region is attenuated in proportion to its instability, so that a contested area of the knowledge base cannot infect a settled one while the contest is unresolved.
Note what this is not. It is not averaging. Two authorities disagree and the system does not produce a midpoint that neither holds and no evidence supports; it records the disagreement as a structural feature, quarantines the region, attenuates its influence, and reports the conflict as the finding. Which is what a competent researcher does, and what a fluent generator, obliged to produce one coherent paragraph, cannot do.
Detecting the conflict at all requires that claims be represented as claims. Text does not contradict. Propositions do. This is the second time the propositional representation earns its cost, and it will not be the last.
VII. The question is data too
Of the eight states, the false premise is the one that punishes responsiveness. “Why did the regulator abolish the reporting requirement in 2019?” — if it did not, the only honest reply is to reject the question. Answering it is not merely wrong; it is the manufacture of a world, complete with plausible reasons, on request. And the more capable the generator, the better the fabricated reasons will be.
The architectural answer follows from a requirement I have already imposed on claims: every proposition entering the system must be classifiable by a decision procedure into one of three epistemic states — verifiable, refutable, or undecidable. Nothing enters unclassified.
Apply that to the prompt. A question carries presuppositions, and presuppositions are claims. Route them through the same admissibility test as anything else. A presupposition that is refutable against the belief base is refuted, and the question is answered by rejecting it rather than by satisfying it. A presupposition that is undecidable is flagged and held, not silently assumed as the ground of everything that follows.
This inverts the default, and the inversion is the whole of it. A generator treats the prompt as authoritative context to be continued — that is what a context window is. An epistemic core treats the prompt as an assertion by an external source of unknown reliability: data, carrying a provenance tag like any other, subject to the same evaluation as a journal article or a sensor reading.
It also explains a failure mode usually filed elsewhere. Leading prompts move a model’s confidence because “explain why X causes Y” supplies a premise the system never tested and then rewards fluent continuation from it. Under this architecture the causal claim is a proposition requiring its own admissibility check before it can serve as a premise, and if it sits below the premise threshold, nothing may be built on it. The prompt loses its power to smuggle. Prompt injection and the leading-question effect are normally treated as adversarial-robustness problems to be patched; they are more accurately described as artefacts of a system that has nowhere to put a presupposition.
VIII. Decay: the answer to “was true once”
Of the eight states, staleness is the one that no output filter can catch, because a stale claim is well-formed, well-sourced and wrong.
The mechanism is decay. Belief in a proposition, absent reinforcement, degrades exponentially in the elapsed time since its last reinforcement, with a decay constant set per domain. Fast-moving empirical fields decay quickly; mathematics and logic decay at negligible rates. The information-theoretic statement is cleaner than the probabilistic one: in the absence of new evidence, the entropy of the belief rises monotonically. Uncertainty never decreases without evidential intervention. That is the whole principle, and it is the exact inverse of how a trained model behaves, since a model’s confidence in a fact learned in pre-training is entirely insensitive to how long ago the fact was true.
Decay is not irreversible. A reinforcement event — a replication, a high-confidence citation, a re-derivation — resets the clock and re-establishes the posterior through ordinary conditioning. Claims that fall below a minimum trust threshold are marked stale and excluded from active inference until revalidated, rather than being deleted; the record persists, and so does the ability to explain why the system stopped relying on it.
This gives the system something no language model has: an answer to “how old is this belief, and does that matter?” The domain-specific decay constant is the encoding of the second half of that question. An atomic mass and the holder of an office are not the same kind of fact, and a system that is equally confident about both is committing a category error at scale.
IX. Verification proportional to consequence
Uniform verification is unaffordable, so in practice it is abandoned, and the abandonment is uniform too. The rule has to be decision-theoretic: weigh the cost of checking against the cost of being wrong.
The architecture encodes this directly. An epistemic loss function assigns distinct penalties to the two error types — believing what is false, and failing to believe what is true — and these need not be symmetric. Each domain carries a risk profile derived from the average of those penalties across its claims. Belief propagation along an edge into a domain is attenuated by that domain’s risk scalar, so that inference into high-consequence territory is intrinsically more conservative than inference into low-consequence territory, without anyone having to remember to be careful.
Above a risk threshold — medical, engineering, safety-critical — the constraint hardens into an admissibility rule: a claim may enter only if it comes from an authoritative source and has been independently replicated or derived from axioms. Not “should be checked”. May not enter.
And where the cost of a false positive greatly exceeds the cost of a false negative, belief updating is biased conservatively until replication or corroboration reduces the expected loss below tolerance. The system’s willingness to commit is a function of what commitment would cost if wrong — which is how any competent professional already operates, and which no current model can represent, because a model’s confidence is a property of the sentence rather than of the stakes attached to it.
X. Chains: the bound that needs no independence assumption
I refused, in the previous essay, to assert that errors compound multiplicatively across reasoning steps. That claim requires the steps to be independent, and independence is an assumption, and untested assumptions are defects. I am not going to smuggle one in to make an argument prettier.
The architecture does not need it. Confidence propagation along a dependency chain is governed by a constraint rather than a product: credence in the antecedents bounds the credence attainable in the conclusion. A conclusion cannot be held more firmly than its weakest necessary support, and inferential uncertainty in the step itself can only lower the ceiling further.
Figure 2. A chain is only as good as its worst link. Bars show credence at each step of a five-step argument; the green line is the propagated ceiling. The conclusion is capped at the weakest antecedent — here 0.61 — regardless of how confident the last three steps were on their own. The red marker is what fluency reports: confidence re-derived from the polish of the final sentence rather than aggregated from anything. Schematic. The ceiling is a bound, not a product, and requires no assumption that the steps are independent.
This is a weaker claim than multiplicative compounding and a more defensible one. It says only that a chain cannot be stronger than its weakest necessary link, which is true whether or not the links are correlated. Where dependencies are positively correlated the true credence may sit well below the ceiling, and the bound remains valid; it is a bound, and bounds do not require you to know the joint distribution.
The complement runs backwards. Contradiction or downgrading at any node triggers recursive re-weighting of everything upstream that depended on it. And the graph is checked for cycles, so that a belief set in which mutual reinforcement masks the absence of any external justification is flagged rather than mistaken for corroboration. Circular support looks exactly like strong support from inside a system with no provenance; from outside, with the justification graph visible, it is obvious.
What this delivers is the property I said was missing: confidence that survives multi-step reasoning, because it is computed over the structure of the reasoning rather than generated afresh at the end.
XI. Active information seeking, and what it should reward
Silence is not enough. “I do not know” leaves the user where they started; “I do not know, and here is the fact that would settle it” is a different product.
The mechanism is to reward learning rather than performance. In the multi-agent formulation, the information gain from an observation is the Kullback–Leibler divergence of the posterior from the prior — a scalar measuring how much the observation actually moved the belief. Agents are re-weighted by a strictly increasing function of that quantity, so an agent whose beliefs never move on any evidence accrues no strength regardless of how well it scores, and its confidence decays under stagnation. What is rewarded is not being right; it is updating.
Set that against how these systems are usually trained, and the contrast is the whole argument. A model penalised for saying “I do not know” learns to guess fluently, because fluent guesses score better than honest silence with any evaluator grading on satisfaction. A system rewarded for information gain learns to seek the observation that would change its mind. Those two objectives produce different machines, and the difference is not a matter of degree.
Alongside this sits epistemic triage. Claims are ranked not only by credence but by their expected contribution to downstream truth — a truth-promotion score measuring the marginal effect on the beliefs a claim influences. High-certainty, low-consequence claims are de-emphasised. Claims that are currently marginal but would resolve a great deal if settled are flagged for targeted re-evaluation. The system is thereby able to say not merely what it does not know, but what it most needs to find out — which is the operational test of whether the uncertainty is real. A system that can name the decisive missing evidence has a model of its own ignorance. A system that shrugs has learned a phrase.
XII. One query, end to end
Abstraction earns nothing until it is walked through. Take a question of the kind that matters: is our indemnity clause enforceable against a foreign subsidiary in this jurisdiction?
The query is parsed into propositions with contextual signatures attached — jurisdiction, entity type, the date at which the question is asked. Its presuppositions are checked: the clause exists in the form described, the subsidiary is constituted as stated. Anything refutable is refuted before the substantive question is touched.
Retrieval follows, and every returned record must satisfy the metadata schema and verify against its hash before it is admitted. Priors are assigned from the venue, the authority level of the source and the methodological character of the material, not from how often the proposition appears in a corpus. Evidence is then applied by type — a controlling decision, a persuasive one, an academic commentary and a contradicting judgment do not contribute the same likelihood, and are not permitted to.
The contradiction graph is consulted, and here the courts are split. The system does not average two lines of authority into a synthetic position no court has taken. It records the contradiction edge, quarantines the region, attenuates propagation out of it, and carries the split forward as a finding rather than a difficulty.
Decay is applied. A 2011 authority in an area that has moved since decays according to the domain constant; unless a reinforcement event — a subsequent affirmation, a replication of the reasoning in a later case — has reset the clock, its contribution is discounted by age rather than treated as though it were decided yesterday.
The domain risk profile is high: the cost of a confident wrong answer here is a liability, not an embarrassment. The hardened admissibility rule therefore applies, and no claim enters the chain without an authoritative source and independent corroboration.
Finally the chain is assembled, and the conclusion is capped at its weakest necessary link — say 0.68, set by the unresolved split. That sits below the assertion threshold. So the system does not assert.
What it returns instead is: the finding, at a stated credence; the location of the split and both lines of authority; the specific appellate question whose resolution would settle it; and what the answer becomes under each resolution. Not an answer. Something more useful than an answer, which is an accurate account of why there is not yet one.
A generator given the same query returns a fluent, confident, correctly-cited paragraph, and no indication whatever that any of the above was in play.
XIII. Where it binds
An architecture that cannot fail is not an architecture. Three constraints are load-bearing, and I will not pretend otherwise.
The truth signal. The multi-agent formulation is built around an external evaluation functional — a truth oracle — that is exogenous, time-invariant, immune to manipulation by the agents, and consistent with a proper scoring rule. Agents are selected on their expected loss against it. That construction is what makes convergence provable, and it is also its sharpest limit, because the oracle is a stipulation. Where the ground truth exists and is cheap — arithmetic, verified experimental outcomes, formally derivable results — the selection pressure is real. Where the ground truth is contested, delayed by decades, or simply unavailable, there is no oracle to select against, and the guarantee thins to nothing.
This is the same problem I raised about calibration measurement generally: verification is cheapest where it matters least and scarcest where it matters most. The architecture does not dissolve it. What it does is refuse to disguise it — replication status, contradiction edges and decay are all visible, so a claim resting on thin verification is legible as such rather than presented in the same voice as a settled one. Making the deficiency inspectable is not the same as solving it, and I am not going to claim otherwise.
The independence of evidence. The cumulative update multiplies likelihood contributions across evidence items, and that step is valid under an explicit conditional-independence condition: that each item’s likelihood given the hypothesis is unaffected by the others. In the scientific corpus that condition frequently fails. Citations cluster. Replications share methods, instruments, reference datasets and sometimes personnel. Two studies from the same lineage are not two independent observations, and multiplying them as though they were overstates the posterior — the exact mechanism by which a field can become collectively confident on a narrow evidential base.
Some countermeasures are in the architecture already: a citation entropy penalty that discounts belief accumulated from a redundant cluster of venues or authors, and an independence requirement restricting the high-confidence tier of replication scoring to replications with no overlapping authors, funding sources or institutions. These are partial. They discount correlated evidence; they do not model the correlation. The correct repair is a joint likelihood with an explicit dependency structure over evidence sources rather than a product over items with an independence proviso attached. Until that is built, the independence condition is an assumption stated rather than a condition demonstrated, and it belongs on the defect list.
The epistemic/aleatoric split. Of the nine requirements, this one has no mechanism. Nowhere is uncertainty decomposed into the part that is reducible by going and finding out and the part that is irreducible because the world has not decided. The distinction is operational, not philosophical: it determines whether searching helps. Sixty per cent because the document exists and has not been read licenses one action; sixty per cent because the coin has not landed licenses the opposite.
Adjacent machinery exists — information gain measures how much an observation would move the belief, and decay tracks the growth of entropy in the absence of evidence — but neither is the split. A single scalar credence cannot carry it. The representation needs two channels, and it currently has one. That is a gap, and naming it is more useful than a paragraph explaining why it nearly counts.
There is a fourth tension worth stating, which is internal rather than external. The commitment framework demands deductive closure: what follows from what you accept, you accept. Taken literally this is logical omniscience, and it is computationally impossible for any real system. The resolution is a locality constraint — belief updates are a function only of local evidence and semantically adjacent claims, propagating along bounded dependency paths — which preserves decidability at the price of the closure ideal. The two commitments pull against each other. The honest description is that closure is a normative target and locality is the operating discipline, and the gap between them is where a real implementation will spend most of its engineering.
XIV. What this costs, and why it is worth paying
Everything above is slower than sampling. Structured claim extraction, provenance verification, contradiction checking across a graph, decay recomputation, risk-weighted propagation and justification logging are all expenditure that a text generator does not incur. The output will be less impressive. It will decline to answer things. It will attach conditions.
The compensating fact is the one from the previous essay: a system that is right less often and tells you which is worth more than a system that is right more often in a uniform voice, because the second one moved the verification work downstream and disguised the move. Every property described here is an instrument for locating the work rather than hiding it. Claim-level tagging tells you what kind of statement you have. Provenance tells you where it came from. Confidence tiers tell you what the system was willing to reason from and what it was willing to say. Contradiction edges tell you where the field is unsettled. Decay tells you how old the belief is. Risk weighting tells you how hard it looked before committing, and why.
None of that makes the machine right. It makes the machine’s rightness checkable, which is the only property that has ever distinguished knowledge from confident assertion — in machines, and in the institutions that have been failing at it for considerably longer.
The useful system is not the one that knows the most. It is the one whose reported uncertainty remains predictively meaningful in domains it was not tuned for, which knows when further evidence is required, which can name what that evidence would be, and which does not convert not-knowing into fluent invention.
We can specify that system. Two of its nine requirements remain open. The specification is worth more than the fluency it gives up.
Classification
Sections II–XII are specification: mechanisms defined in the cited papers, restated. Section X’s propagation bound is deductive and deliberately weaker than multiplicative compounding, which would require an independence assumption I decline to make. Section XIII is a defect register: the exogenous truth oracle is a stipulation, not a demonstration; the conditional-independence condition on evidence aggregation is asserted rather than established and is empirically doubtful for citation and replication data; the epistemic/aleatoric decomposition is absent, not approximated. The figures are schematic and report no measurement. Nothing here is an implementation result — no system described has been built and evaluated, and no performance claim is made or implied.
Sources
-
Wright, C. S. (2025). Bayesian Epistemology with Weighted Authority: A Formal Architecture for Truth-Promoting Autonomous Scientific Reasoning. arXiv:2506.16015. — claim representation, prior construction, evidence-typed updating, contradiction graph and quarantine, decay, truth-promotion score, risk-aware weighting.
-
Wright, C. S. (2025). Beyond Prediction — Structuring Epistemic Integrity in Artificial Reasoning Systems. arXiv:2506.17331. — propositional commitment, confidence stratification and assertion rights, contradiction as system failure, statistical engine / epistemic core separation, confidence propagation, disclosability.
-
Wright, C. S. (2025). Bayesian Evolutionary Swarm Architecture: A Formal Epistemic System Grounded in Truth-Based Competition. arXiv:2506.19191. — truth oracle as exogenous evaluation functional, strictly proper scoring utility, KL information gain and confidence reweighting, entropy regularisation, evolutionary selection on epistemic fitness.