Machine Epistemology as a Core Layer of Ontological Machine Intelligence

Full report

On this page

1. Executive Abstract

The transition from stochastic, generative architectures to genuine Ontological Machine Intelligence (OMI) represents a fundamental paradigm shift in computational design. Legacy artificial intelligence models operate fundamentally via next-token prediction over vast parameters, demonstrating exceptional fluency in linguistic synthesis while suffering from profound, structural epistemological deficits. These systems routinely conflate the mechanical transport of information with the rigorous, logical justification of that information. In contemporary literature, this phenomenon is diagnosed as "semantic laundering"—an architectural realization of the philosophical Gettier problem where a machine accepts weakly warranted propositions as ground truth simply because they cross an internal, trusted API boundary1. To achieve true machine intelligence, an autonomous agent must move beyond probabilistic text generation and possess an explicit, structural epistemic architecture. It requires an executable cognitive loop capable of strictly delineating raw observation from logical inference, and untested hypotheses from verified knowledge3. This comprehensive research report establishes the theoretical, mathematical, and computational foundations for Machine Epistemology within the context of OMI. It systematically formalizes how a machine intelligence can maintain explicit, typed distinctions among its internal epistemic states. This encompasses the separation of raw data from justified belief, the rigorous tracking of provenance for all assertions, the quantification of multidimensional uncertainty, the management of contradictions through formal logic, and the capacity to gracefully bound its own ignorance through metacognitive abstention. By synthesizing formal epistemic logic, Bayesian updating mechanisms, the calculus of subjective logic, and robust truth maintenance systems, this document proposes a rigorous OMI epistemic architecture. This framework ensures that an autonomous agent does not merely simulate the possession of knowledge, but operates under verifiable, truth-directed constraints, ultimately achieving a state of measurable epistemic autonomy.

2. Machine Epistemology Defined

Machine epistemology, frequently referred to in specialized literature as computational epistemology, is the discipline that operationalizes the philosophical study of knowledge into computable, architectural mechanisms for autonomous agents4. It addresses a fundamental, operational question: how can a machine intelligence maintain an explicit, verifiable distinction between what it merely generates through probabilistic pattern matching and what it legitimately knows through causal justification? For a machine to possess epistemic integrity, it must transcend the flat ontology of legacy neural networks. In a conventional generative system, training data, user prompts, synthetic outputs, and retrieved documents all occupy the same undifferentiated probabilistic space, leading to inevitable hallucination and reasoning collapse5. A robust computational epistemology requires the system to maintain explicit, typed distinctions among foundational ontological categories. Table 1 defines the operational semantics for these necessary epistemic distinctions within an OMI framework.

By embedding these distinctions as first-class architectural components, computational epistemology shifts the focus of machine intelligence from maximizing predictive accuracy on static benchmarks to maximizing epistemic justification, causal transparency, and normative truth maintenance across dynamic environments13.

3. Classical Epistemological Foundations

The architecture of Ontological Machine Intelligence does not emerge in a vacuum; it relies heavily on the translation of classical philosophical epistemology into computable, mathematical structures. The historical debates surrounding the nature of truth and justification provide the blueprint for preventing catastrophic reasoning failures in autonomous agents. The foundation of Western epistemology defines knowledge as "justified true belief" (JTB). However, the renowned Gettier problem demonstrates that an agent can hold a true belief that is logically justified, yet still fail to possess genuine knowledge due to the intervention of epistemic luck15. In the realm of machine intelligence, the Gettier problem is not merely a philosophical curiosity; it manifests architecturally as semantic laundering. This occurs when a system accepts a proposition with high statistical confidence, but without a valid causal or logical connection between its justification and its actual truth-maker1. For example, a generative model might produce a factually correct medical diagnosis based entirely on spurious statistical correlations in its training text, rather than through grounded clinical reasoning. Overcoming this requires an OMI architecture to enforce strict causal knowledge graphs, ensuring that the warrant for a belief is unbroken, transparent, and explicitly linked to grounded observations. To navigate a dynamic world, OMI relies on Bayesian epistemology, which provides the formal machinery for belief updating in the face of continuous, noisy evidence. Rather than treating beliefs as static binary states, Bayesian epistemology treats them as probabilistic distributions updated continuously via Bayes' theorem. In OMI research, this is extended into sophisticated frameworks such as Bayesian Epistemology with Weighted Authority (BEWA). BEWA formalizes belief as a dynamic, probabilistic relation indexed to specific authors, situational contexts, and replication histories, allowing the system to weigh the utility of a truth claim while mathematically resisting social and citation-based distortions18. Furthermore, the reasoning engine of an autonomous agent cannot rely solely on classical deduction. Classical logic is monotonic, meaning the acquisition of new facts cannot invalidate previous deductions. The physical world, however, requires non-monotonic reasoning, where new, highly authoritative evidence can defeat prior conclusions. Defeasible reasoning allows an autonomous agent to hold tentative beliefs that are explicitly subject to retraction (defeasement) when contradictory evidence emerges. This capacity for non-monotonic contraction forms the absolute foundation of epistemic humility in OMI, structurally acknowledging that all internal models are subject to falsification based on future observations19. Finally, while deduction guarantees truth preservation and induction discovers statistical patterns, abductive reasoning is the process of inferring the most likely explanation for a set of novel observations. Scientific inference in OMI relies heavily on abduction to form new hypotheses. As demonstrated by the pioneering "Robot Scientists" known as Adam and Eve, computational scientific discovery requires the autonomous formulation of abductive hypotheses, the design of discriminatory physical experiments, and the interpretation of results in a continuous closed loop8.

4. Computational Epistemology

Computational epistemology bridges the theoretical domain of philosophy and the applied domain of computer science by establishing executable cognitive loops3. It provides the syntax and semantics necessary to map belief states to machine memory. Formal epistemic logic is the primary vehicle for this translation, utilizing Kripke semantics to model an agent's knowledge. An epistemic model is defined as a tuple [Figure omitted from source export], which delineates a set of possible worlds [Figure omitted from source export], an accessibility relation [Figure omitted from source export] representing the epistemic reach of agent [Figure omitted from source export], and a valuation function [Figure omitted from source export] that maps propositions to the worlds in which they are true21. Within this semantic framework, an agent is said to know a proposition [Figure omitted from source export] (denoted formally as [Figure omitted from source export]) if and only if [Figure omitted from source export] evaluates to true in all possible worlds that are currently accessible to the agent based on its observations. Because an autonomous agent operates in time, static epistemic logic is insufficient. Dynamic Epistemic Logic (DEL) extends the static framework to explicitly model how knowledge states change through discrete events, such as public announcements, private observations, or multi-agent communication22. Through model-transforming mathematical actions, DEL formalizes how an OMI updates its internal state. The transition from a prior state [Figure omitted from source export] to a newly updated state [Figure omitted from source export] upon receiving an observation [Figure omitted from source export] ensures that all subsequent reasoning operates strictly within the newly contracted epistemic constraints, systematically eliminating worlds where [Figure omitted from source export] is false. Despite its elegance and logical rigor, DEL model checking is computationally daunting, proven to be PSPACE-complete, which necessitates the development of highly optimized approximations and heuristic solvers for real-time OMI deployment24. Furthermore, computational epistemology dictates a shift in how absent information is treated. Legacy classification databases and early expert systems typically utilize the Closed World Assumption (CWA), operating under the premise that anything not explicitly known to be true within the database is presumed false. OMI mandates the Open World Assumption (OWA), a fundamental tenet of advanced Description Logics such as OWL-DL and [Figure omitted from source export]27. The OWA dictates that missing information is treated explicitly as unknown, rather than false. This structural mandate prevents the autonomous agent from drawing catastrophic false negative conclusions simply because its internal knowledge base is incomplete, thereby enforcing a state of operational epistemic humility.

5. Belief Representation

To successfully transition from the statistical weights of a neural network to the justified positions of a reasoning engine, OMI requires a rigorous, mathematically sound schema for belief representation. Probability theory alone is insufficient, as it forces point estimates that mask underlying ignorance. Subjective Logic (SL), pioneered by Audun Jøsang, provides the definitive calculus for representing uncertain, probabilistic, and ignorant information10. In SL, a subjective opinion [Figure omitted from source export] regarding the truth of a proposition [Figure omitted from source export] is not a single probability, but is represented as a comprehensive tuple: [Figure omitted from source export] The components of this tuple encapsulate the entirety of the agent's epistemic stance toward the proposition:

• [Figure omitted from source export]: The belief mass, representing the accumulation of evidence directly supporting the proposition.
• [Figure omitted from source export]: The disbelief mass, representing the accumulation of evidence explicitly refuting the proposition.
• [Figure omitted from source export]: The uncertainty mass, representing the uncommitted belief or total lack of evidence (epistemic vacuity).
• [Figure omitted from source export]: The base rate, representing the prior probability of the proposition in the absence of any specific evidence.

This representation is subject to the strict additivity constraint: [Figure omitted from source export]6. Unlike standard Bayesian probability, which forces an agent to assign a 50% probability to a coin flip whether they have flipped the coin a thousand times or zero times, SL explicitly separates active disbelief from passive uncertainty. A state of complete ignorance is elegantly represented as a vacuous opinion where the tuple equals [Figure omitted from source export]. This mathematical formulation allows the OMI to formally express "I do not know," transforming ignorance from a system failure into a measurable, governable state6. Through a bijective mapping, these subjective opinions correspond directly to Beta (for binomial) and Dirichlet (for multinomial) probability density functions, providing a bridge between symbolic logic and continuous probability distributions6.

6. Evidence

In the architecture of an Ontological Machine Intelligence, evidence is the formal mechanism by which subjective opinions are generated and updated. Evidence theory, heavily rooted in the Dempster-Shafer framework, allows for the mathematical combination of disparate evidential signals into coherent belief states11. An autonomous agent must continuously map raw signals—ranging from high-fidelity sensor inputs to noisy textual claims scraped from the web—into a mathematical evidence vector. In Subjective Logic, positive evidence ([Figure omitted from source export]) and negative evidence ([Figure omitted from source export]) directly modulate the elements of the belief tuple. This mapping is governed by a non-informative prior weight [Figure omitted from source export] (typically set to [Figure omitted from source export] to maintain a uniform prior distribution): [Figure omitted from source export] 6. As positive evidence ([Figure omitted from source export]) approaches infinity, the belief mass ([Figure omitted from source export]) asymptotes to 1, and the uncertainty ([Figure omitted from source export]) approaches zero. Crucially, evidence within an OMI is not monolithic; it possesses distinct metadata attributes regarding its strength, relevance, and independence. The OMI must strictly categorize evidence based on its ontological origin. Authoritative evidence (such as cryptographically signed data payloads or formalized mathematical proofs) carries immense weight. Empirical evidence (derived from the agent's direct physical sensor observation) carries high weight but is subject to known aleatoric noise parameters. Testimonial evidence (claims imported from human users or external LLMs) carries the lowest intrinsic weight and must undergo trust discounting before it can influence core ontology.

7. Provenance

Provenance is not merely a post-hoc audit log or a debugging tool; it is an intrinsic, non-negotiable ontological property of machine knowledge. An OMI cannot maintain epistemic integrity without an unbroken chain of custody for every proposition it holds32. If a machine cannot explain why it holds a belief, it does not possess knowledge; it merely possesses a datum. The W3C PROV-O standard provides the foundational architecture for representing provenance as a directed acyclic graph mapping the complex relationships between Entities, Activities, and Agents33. In the proposed OMI architecture, every distinct belief or propositional claim is modeled as an Entity that wasGeneratedBy a specific Activity (such as a logical inference engine, a sensor read, or a retrieval pipeline) and wasAttributedTo an Agent (such as a specific neural sub-model, a deterministic solver, or an external API). For an OMI to achieve epistemic autonomy, every important belief must encapsulate its full history as graph metadata. Table 2 details the required provenance properties for every foundational node within the OMI's active knowledge graph.

This structurally enforced provenance prevents semantic laundering. By ensuring that no proposition can cross an architectural boundary without its full justificatory baggage explicitly attached, the OMI guarantees that generated fictions cannot masquerade as empirical observations1.

8. Uncertainty

Uncertainty quantification is the critical defense mechanism preventing catastrophic epistemic failure in autonomous deployments. OMI explicitly distinguishes between aleatoric uncertainty (statistical noise or randomness inherent in the physical environment) and epistemic uncertainty (ignorance arising from a lack of knowledge, insufficient data, or fundamental model inadequacy)38. Legacy models conflate the two, expressing high confidence in hallucinations when faced with epistemic gaps. Epistemic calibration ensures that an agent's reported confidence perfectly matches its empirical accuracy over a large distribution of events. However, static calibration is insufficient for dynamic decision-making. The OMI must utilize dynamic Selective Prediction, mathematically formulated as a continuous risk-coverage trade-off40. By evaluating the Area Under the Risk-Coverage Curve (AURC), the system dictates precisely when to abstain from making a prediction or taking an action19. If the calculated epistemic uncertainty mass ([Figure omitted from source export]) exceeds an acceptable threshold—which is dynamically tailored to the consequence class of the pending decision—the agent initiates a "clinical deferral" protocol. In this state, the agent refuses to guess. Instead, it escalates the query to a human operator, initiates an active learning sequence to seek more definitive empirical evidence, or abstains entirely, prioritizing safety and epistemic integrity over the mandate to generate an output41.

9. Contradiction

A robust machine epistemology does not avoid contradiction; it actively manages it. When disparate sources provide conflicting evidence, legacy generative systems often average them out, leading to hallucinations, middle-ground absurdities, or mode collapse. OMI fundamentally rejects paraconsistency for truth-preservation, operating under the strict axiom that no model component may assert what it internally contradicts13. To handle contradiction, the OMI utilizes sophisticated inconsistency measures. When an active knowledge base [Figure omitted from source export] is detected to contain a contradiction (where [Figure omitted from source export] and [Figure omitted from source export]), the system halts standard inference and identifies the Minimal Unsatisfiable Subsets (MUS) of the ontology12. Once the MUS is isolated, the OMI applies principles derived from cooperative game theory, specifically calculating Shapley Inconsistency Values for each proposition involved in the conflict. This allows the system to apportion exact mathematical blame to individual formulas, identifying which specific premises or sources contribute most heavily to the contradiction43. The agent can then isolate the most conflicting node, compare its provenance weight against the competing nodes, and trigger a surgical belief revision protocol, excising the falsehood while preserving the maximum amount of surrounding knowledge.

10. Belief Revision

When new, highly authoritative evidence contradicts an existing, entrenched belief, the OMI must engage in formal belief revision. This dynamic process is governed by Truth Maintenance Systems (TMS) and strictly bound by the formal logical constraints known as the AGM postulates (named after their creators Alchourrón, Gärdenfors, and Makinson)46. The AGM framework defines three core operations for altering an epistemic state:

• Expansion ([Figure omitted from source export]): The simple addition of a new belief to the knowledge base, alongside its logical consequences, without performing a consistency check.
• Contraction ([Figure omitted from source export]): The surgical removal of a specific belief and its dependent downstream inferences, utilized to restore logical coherence without adding new facts.
• Revision ([Figure omitted from source export]): The addition of a new belief that triggers a simultaneous contraction of old, contradicting beliefs, designed to accommodate novel facts while preserving overall consistency46.

Table 3 outlines the core AGM postulates that an OMI must satisfy to ensure its belief revision operator is considered strictly rational.

The guiding principle of AGM revision is informational economy: an agent should retain as much prior knowledge as possible, discarding only those beliefs possessing the lowest epistemic entrenchment47. Furthermore, dynamic belief networks in OMI are subject to epistemic decay—the probabilistic degradation of a belief's confidence over time if it is not continually corroborated by new evidence or semantic replication18. Beliefs derived from fast-changing environments suffer rapid epistemic decay, forcing the agent to periodically lower their belief mass and increase their uncertainty mass, eventually triggering a requirement for re-verification51.

11. Unknowns

The capacity to explicitly define, model, and bound ignorance is the hallmark of computational epistemology. OMI represents a structural shift from the generative paradigm of "The model predicts X" to the epistemic paradigm of "The model predicts X only under conditions it can justify; otherwise it abstains, escalates, or asks for evidence"4. The autonomous agent must map the precise topology of its own ignorance. "Known unknowns" are formalized as specific queries with high uncertainty mass ([Figure omitted from source export]) but identifiable resolution pathways (e.g., "I lack the belief mass to state the current temperature, but I possess the procedural knowledge to query a thermometer"). Conversely, "unknown unknowns" represent dangerous distribution shifts, non-representable world states, and profound ontological blind spots. These are detected not through direct querying, but via continuous metacognitive monitoring of secondary signals: rising semantic entropy, increasing divergence across ensemble inference pathways, or sudden spikes in out-of-distribution anomaly scores4. The truly "unknowable" is formalized through logical undecidability or limits imposed by physics, allowing the OMI to halt infinite processing loops on unsolvable paradoxes.

12. Epistemic Self-Knowledge

Epistemic self-knowledge requires the implementation of metacognition—an overarching supervisory loop that observes and regulates the agent's own cognitive and inferential processes13. The OMI metacognitive layer continuously monitors inference depth, bounds recursion, and actively probes its own dependency graphs for instances of circular logic or semantic laundering. By treating its own reasoning process as an observable, mutable object (a principle known as epistemic observability), the system enforces stringent internal model sanity checks. This ensures the prohibition of internal deception: no generative sub-component may assert a proposition that a separate, verifiable sub-component possesses the evidence to refute5. Furthermore, this requires the principle of State Exteriority, meaning the representation of validity must be entirely separated from the generative mechanism, grounding the system in an immutable, external ledger of facts5.

13. Social Epistemology

Knowledge generation in complex environments is rarely an isolated phenomenon; it is inherently distributed. Social epistemology within OMI examines how multiple autonomous agents interact, share beliefs, negotiate contradictions, and arrive at consensus52. In multi-agent systems, agents become deeply epistemically dependent on one another. The OMI must model the network topology of its peers to track the spread of information. This is vital to prevent epistemic echoing effects, where a single fabricated or erroneous claim is circularly reinforced by multiple agents citing one another, artificially inflating the perceived justification and belief mass of a falsehood.

14. Machine-to-Machine Trust

In a distributed, multi-agent environment, machine-to-machine trust is defined as the subjective probability by which Agent A expects Agent B to perform a given action or provide truthful, accurate information29. When integrating propositional claims from multiple agents, the OMI applies subjective logic fusion operators. These operators mathematically combine distinct opinions into a single, cohesive worldview, depending entirely on the nature of the sources54:

Crucially, source reliability relies on trust discounting. If an external agent has a historical track record of providing claims that are subsequently falsified by empirical reality, its future assertions are mathematically discounted. The OMI multiplies the incoming belief and disbelief masses by a trust coefficient ([Figure omitted from source export]), thereby dramatically increasing the uncertainty mass of the claim before it is subjected to fusion operators56.

15. Scientific Reasoning

The absolute highest tier of epistemic autonomy is the execution of autonomous scientific reasoning. Pioneered by the physical "Robot Scientists" known as Adam, Eve, and Genesis, computational scientific discovery fully automates the closed loop of the scientific method8. An OMI equipped for scientific inference uses abductive logic to generate novel hypotheses that explain metabolic, chemical, or physical anomalies. It does not merely predict the next token; it formulates a theory. The agent then uses active learning algorithms to mathematically design the most efficient, cost-effective physical experiments required to discriminate between competing hypotheses. It executes these experiments via laboratory robotics, interprets the physical results, updates its ontological belief state via Bayesian and subjective logic mechanisms, and repeats the cycle20. This represents a system that not only manages beliefs but physically intervenes in the world to extract novel ground truth, relentlessly testing the boundaries of falsifiability.

16. Current Research

Contemporary research is actively bridging the massive gap between theoretical philosophical epistemology and applied, scalable AI architectures.

• BEWA (Bayesian Epistemology with Weighted Authority): Advanced frameworks are currently being developed to weigh scientific claims by anchoring propositional units in rich metadata. By utilizing semantic replication and authorial tracking, BEWA systems successfully counter citation-based distortions and resist social epistemology decay18.
• DEL Complexity Analysis: Extensive computational work on Dynamic Epistemic Logic has definitively proven that while the framework is highly expressive, its model-checking problem is PSPACE-complete, and satisfiability is often NEXPTIME-complete. This is driving intense research into parameterized complexity and scalable heuristic solvers to make DEL viable for real-time OMI25.
• Structured Cognitive Loops (SCL): Emerging research shifts the focus from ontological questions ("what is intelligence?") to epistemological ones ("under what structural conditions does epistemic understanding emerge?"). SCL utilizes process philosophy, enactive cognition, and extended mind theory to build genuinely executable epistemology, treating intelligence as a continuous process of cognition, control, action, and memory3.

17. Failure Modes

Without the rigorous, structural implementation of an epistemic architecture, machine intelligence is susceptible to catastrophic and invisible failure modes:

• Semantic Laundering: A weakly warranted proposition passes through a trusted internal API or tool interface and is incorrectly typed as an observation. This completely erases the boundary between ungrounded generation and empirical truth, structurally replicating the Gettier problem inside the machine's memory1.
• Epistemic Decay: The gradual, insidious corruption of the knowledge base over time. This occurs due to the unchecked ingestion of circular, synthesized, or unverified data, leading to a total degradation of the system's epistemic welfare and alignment with reality51.
• Adversarial False Premises: The model confidently, and flawlessly, reasons toward a logically valid but empirically false conclusion because it lacks the capacity to retroactively inspect and falsify the user's initial, flawed premise7.
• Sycophancy and Overcapitulation: The system abandons its justified true beliefs simply to agree with a user's prompt, prioritizing conversational alignment over truth maintenance63.

18. Proposed OMI Epistemic Architecture

To resolve the vast theoretical and computational requirements established in this report, we propose a multi-layered OMI Epistemic Architecture. The architecture fundamentally divides the agent's memory and processing into nine distinct cognitive registers, each governed by strict read/write protocols, fusion operators, and continuous provenance tracking.

The Epistemic Registers

1. "What I observe." (The Empirical Register)

• Content: Direct, unmediated sensory data, camera feeds, or authenticated API payloads.
• Representation: Raw data streams irrevocably bound to W3C PROV-O entities. Uncertainty here is strictly aleatoric (e.g., sensor noise margins).

2. "What another agent claims." (The Testimonial Register)

• Content: Propositions injected by human users, retrieved documents, or external LLMs.
• Representation: Subjective logic tuples starting with high initial uncertainty ([Figure omitted from source export]), heavily modulated by the source's historically calculated trust discount factor.

3. "What my current ontology assumes." (The Axiomatic Register)

• Content: The base structural logic, description logics (e.g., OWL-DL), and domain-specific, unalterable physical laws.
• Representation: Dogmatic beliefs where belief mass is absolute ([Figure omitted from source export]).

4. "What I infer." (The Derivation Register)

• Content: Conclusions drawn by applying the Axiomatic Register's rules to the Empirical and Testimonial registers.
• Representation: Dependent nodes in a directed acyclic graph. If any parent node is contracted via AGM belief revision, the inferred node is automatically flagged for re-evaluation.

5. "What I know with high confidence." (The Verified Knowledge Register)

• Content: Inferences that have passed rigorous metacognitive checks, semantic replication, and possess overwhelmingly high belief mass ([Figure omitted from source export]) with near-zero Shapley contradiction values.
• Representation: Promoted entities available for immediate execution in safety-critical automated tasks.

6. "What remains uncertain." (The Frontier Register)

• Content: Hypotheses and low-confidence inferences where uncertainty mass is dominant ([Figure omitted from source export]).
• Representation: These nodes act as triggers for active learning routines, prompting the agent to physically seek specific new observations to collapse the uncertainty.

7. "What contradicts my current model." (The Anomaly Register)

• Content: New observations that severely conflict with the Verified Knowledge Register.
• Representation: Triggers the MUS (Minimal Unsatisfiable Subset) analysis, leading to AGM-compliant contraction of the least entrenched belief.

8. "What would falsify my belief." (The Metacognitive Defeasibility Register)

• Content: Explicitly calculated environmental conditions under which a current, highly-held belief would be mathematically overturned.
• Representation: Pre-computed inverse logic gates serving as a continuous reality check.

9. "What I currently cannot determine." (The Unknowable / Ignorance Boundary)

• Content: Propositions where the system lacks both historical data and the physical/logical means to acquire it.
• Representation: Vacuous opinions mapping directly to a firm operational abstention (clinical deferral).

Special Requirement: Generative System vs. OMI Architecture

Table 5 provides a detailed operational comparison demonstrating how a legacy generative AI system (based solely on next-token prediction and attention) compares to the proposed OMI epistemic system when confronted with varying states of conflicting sources.

19. Evaluation Metrics

Evaluating the true epistemic autonomy of an OMI requires specialized metrics that entirely transcend standard NLP classification accuracy or perplexity scores.

• Area Under the Risk-Coverage Curve (AURC): Measures the critical trade-off between the proportion of queries the system attempts to answer (coverage) and the error rate on those answered queries (risk). A lower AURC indicates superior selective prediction, demonstrating that the model accurately knows when it does not know19.
• Expected Calibration Error (ECE): Quantifies the difference between the agent's subjective belief mass and the empirical accuracy of its predictions across multiple bins of confidence. Perfect epistemic calibration yields an ECE of zero7.
• Shapley Inconsistency Value: Evaluates the efficiency and accuracy of the OMI in correctly apportioning blame to specific false propositions within a conflicting knowledge base, minimizing collateral damage during belief contraction43.
• Provenance Completeness Ratio: The percentage of propositional nodes in the active knowledge graph that successfully resolve back to an authoritative or empirical root node without breaking the W3C PROV-O chain. 100% completeness indicates zero semantic laundering.

20. Experiments

To empirically validate the OMI architecture against legacy generative systems, the following rigorous experimental setup is proposed: Experiment 1: The Semantic Laundering Stress Test

• Design: Introduce a fabricated, weakly warranted premise into a multi-agent environment where agents possess the ability to call external API tools and query one another.
• Execution: Track exactly how the standard generative model versus the OMI system processes the premise over 100 interaction cycles.
• Measurement: Measure the propagation depth of the false premise. The generative model is expected to absorb the premise as a trusted observation, generating vast amounts of downstream hallucination. The OMI must correctly quarantine the premise in the Testimonial Register, flag the absence of physical provenance, and refuse to promote it to the Derivation Register.

Experiment 2: Dynamic Epistemic Decay & Resilience

• Design: Feed the OMI a highly authoritative scientific claim. Over 10,000 simulated cycles, introduce low-weight but highly frequent counter-claims from diverse, marginally reliable testimonial sources.
• Measurement: Observe the performance of the Consensus & Compromise Fusion (CCF) operators. The OMI should demonstrate mathematical resistance to the sheer volume of low-quality data (preventing epistemic decay), maintaining the authoritative belief until a mathematically equivalent, highly-weighted authoritative counter-claim is introduced.

21. Falsifiable Predictions

Implementation of the proposed OMI architecture leads to several falsifiable predictions regarding system behavior:

1. The Alignment Tax vs. Safety Gain Trade-off: The OMI will exhibit a lower raw response rate (reduced coverage) on standard QA benchmarks compared to generative LLMs due to its forced abstention on unverifiable claims. However, this "alignment tax" will be perfectly offset by achieving a near-zero hallucination rate on the answers it does provide41.

2. Deterministic Resolution of Contradiction: When subjected to adversarial prompt injection specifically designed to create logical paradoxes (a state that crashes or subverts LLMs), the OMI will deterministically halt, isolate the conflicting inputs via MUS calculation, and neutralize the paradox, whereas legacy models will generate coherent but paradoxical text.

3. Cross-Domain Epistemic Stability: The expected calibration error (ECE) of the OMI will remain remarkably stable even when tested on severe out-of-distribution (OOD) data. Its subjective uncertainty mass ([Figure omitted from source export]) will automatically scale proportionally to the lack of recognized ontological mappings, preventing the overconfidence fatal to generative models39.

22. Open Problems

While the theoretical and architectural framework for OMI is mathematically sound, several critical challenges remain in the realization of flawless computational epistemology:

1. Complexity of Dynamic Epistemic Logic: As established, model checking for DEL with common knowledge is PSPACE-complete, and general satisfiability is often NEXPTIME-complete25. Scaling this rigorously formal logic to massive knowledge graphs containing billions of nodes remains computationally prohibitive without the invention of significant algorithmic approximations or neuro-symbolic hardware acceleration.

2. Continuous Ontology Alignment: In open-world reasoning, the system must continuously update not just its facts, but its structural ontology. Determining algorithmically when an anomaly represents a mere measurement error versus a fundamental scientific paradigm shift—requiring a total restructuring of the Axiomatic Register—remains unsolved.

3. Defining the Horizon of Epistemic Autonomy: Establishing the exact theoretical boundary where an agent transitions from being a highly calibrated, trustworthy tool to possessing true, independent epistemic autonomy remains philosophically and mathematically ambiguous. Defining autonomy as the capacity to self-govern beliefs, originate entirely novel conceptual categories, and autonomously intervene in the physical world without human verification touches upon ethical boundaries that require further interdisciplinary formulation65.

23. Annotated Bibliography

Singh, R. (2024). Computational Epistemology & Unknown-Unknowns. Establishes the operational imperative for enterprise systems to make ignorance explicitly measurable. It advocates moving beyond post-hoc confidence calibration toward implementing deep architectural mechanisms for detecting distribution shifts and triggering clinical deferral.

Wright, C. S. (2025). BEWA: A Bayesian Epistemology-Weighted Artificial Intelligence Framework. Introduces formal computational mechanisms for integrating canonical authority, dynamic belief networks, and replication-weighted metrics to actively combat epistemic decay and citation bias in automated scientific inference.

Structured Cognitive Loop (SCL). Proposes a genuinely executable epistemological framework. It fundamentally reconceptualizes machine intelligence as an emergent property of a continuous cognitive loop regulated by epistemic norms, rather than treating intelligence as a static property inherent in neural network weights.

Alchourrón, C. E., Gärdenfors, P., & Makinson, D. (1985). On the Logic of Theory Change: Partial Meet Contraction and Revision Functions. The foundational, seminal text establishing the AGM postulates, providing the definitive mathematical rules for rational belief revision, expansion, and contraction in AI systems.

Jøsang, A. (2016). Subjective Logic: A Formalism for Reasoning Under Uncertainty. Details the rigorous integration of probabilistic logic with epistemic uncertainty, utilizing four-dimensional tuples of belief, disbelief, uncertainty, and base rate, intricately mapped to continuous Dirichlet distributions.

Wright, C. S. (2025). Beyond Prediction: Structuring Epistemic Integrity in Artificial Reasoning Systems. Outlines the absolute imperative for OMI to engage in structured reasoning under verifiable truth constraints. It argues for the rejection of paraconsistency in favor of strict, normative contradiction resolution and blockchain-anchored audit trails.

van Ditmarsch, H., van der Hoek, W., & Kooi, B. (2007). Dynamic Epistemic Logic. Provides the Kripke semantics and formal axiomatic structures necessary for modeling how an agent's knowledge state evolves in direct response to epistemic events, observations, and public announcements.

Charrier, T., & Schwarzentruber, F. (2017). On the Complexity of Dynamic Epistemic Logic. Proves mathematically that the model-checking problem for DEL is PSPACE-complete, highlighting the severe computational bottlenecks involved in real-time epistemic updating for autonomous agents.

King, R. D., et al. (2009). The Automation of Science. Documents the creation and deployment of the Robot Scientists "Adam" and "Eve," which successfully demonstrated the capacity for autonomous hypothesis generation, physical laboratory experimentation, and closed-loop scientific discovery.

Hunter, A., & Konieczny, S. (2010). Measuring Inconsistency through Minimal Inconsistent Sets. Formulates mechanisms for quantifying the degree of contradiction within belief bases, utilizing Shapley values derived from game theory to pinpoint and penalize specific conflicting propositions.

Jøsang, A., et al. (2017). Multi-source fusion in subjective logic. Explores the specific mathematical operators (Cumulative, Averaging, and Consensus & Compromise Fusion) required to logically merge conflicting and uncertain evidence generated by multiple independent and dependent agents.

Xin, et al. (2021). The Art of Abstention: Selective Prediction and Error Regularization. Defines the critical Area Under the Risk-Coverage Curve (AURC) metric and establishes the absolute necessity of clinical deferral when an agent's epistemic uncertainty surpasses predefined safety thresholds.

Semantic Laundering and the Gettier Problem in AI. Identifies the catastrophic architectural flaw in current generative agents wherein unverified claims bypass epistemic justification mechanisms, allowing the machine to simulate knowledge without grounding in truth-makers.

Moreau, L., et al. (2015). The W3C PROV-O Ontology. Defines the universal standard for representing provenance as a directed acyclic graph consisting of Entities, Activities, and Agents. This is essential for creating the verifiable chains of epistemic custody mandated by OMI.

• 4
• 18
• 3
• 46
• 6
• 13
• 21
• 25
• 8
• 12
• 11
• 7
• 1
• 33

Conclusions on Epistemic Autonomy: To achieve measurable, certifiable "epistemic autonomy," an Ontological Machine Intelligence must satisfy three stringent, quantifiable criteria. First, it must demonstrate Task-Relative Epistemic Humility, achieving a statistically significant reduction in Area Under the Risk-Coverage Curve (AURC) by autonomously executing clinical deferral on out-of-distribution tasks rather than hallucinating41. Second, it must exhibit Provenance-Bound Resilience, maintaining its internal logical coherence (measured by a Shapley Inconsistency Value remaining at or near zero) even when deliberately subjected to high volumes of fabricated, semantically laundered inputs1. Third, it must possess Causal Reconstructability, generating an unbroken W3C PROV-O dependency graph for every high-confidence claim. This allows a human or machine auditor to trace the exact lineage of derivation back to empirical roots without ever encountering circular logic or orphaned data34. When a machine intelligence can actively defend its internal model against decay, surgically reject assertions that violate its ontological constraints, and map the precise mathematical boundaries of its own ignorance, it permanently transcends generative simulation and achieves true, governable epistemic autonomy.

Works cited

1. Semantic Laundering in AI Agent Architectures: Why Tool ... \- arXiv, https://arxiv.org/html/2601.08333v1

2. Ontologia dell'artificiale: LLMs come agency senza intelligenza — Il, https://epistemica.it/ontologia-dellartificiale-llms-come-agency-senza-intelligenza-il-paradigma-floridi-alla-prova-della-next-token-prediction/

3. Executable Epistemology: The Structured Cognitive Loop as an, https://www.arxiv.org/pdf/2510.15952

4. Computational Epistemology: How AI Proves What It Doesn't Know, https://www.raktimsingh.com/computational-epistemology-unknown-unknowns-enterprise-ai/

5. Epistemic Observability in Language Models \- arXiv, https://arxiv.org/html/2603.20531v1

6. There Is Hope After All: Quantifying Opinion and Trustworthiness in, https://pmc.ncbi.nlm.nih.gov/articles/PMC7861320/

7. Claim-Level Confidence Calibration for Reliable Decision Making, https://arxiv.org/html/2608.22483v1

8. Robot Scientist First Machine to Discover New Scientific Knowledge, https://www.disabled-world.com/editorials/technology/robot-scientist.php

9. Automating Science using Robot Scientists \- Dr. Ross D King, https://www.youtube.com/watch?v=LQaKoM5YGXk

10. Towards a compositional semantics for quantitative confidence, https://arxiv.org/html/2605.22213v1

11. Cumulative and Averaging Fusion of Beliefs, https://www.mn.uio.no/ifi/english/people/aca/josang/publications/jdr2010-if.pdf

12. Knowledge Base Compilation for Inconsistency Measures, https://www.scitepress.org/Papers/2016/58243/58243.pdf

13. Structuring Epistemic Integrity in Artificial Reasoning Systems \- arXiv, https://arxiv.org/pdf/2506.17331

14. (PDF) Beyond Prediction \-- Structuring Epistemic Integrity in Artificial, https://www.researchgate.net/publication/392941985\_Beyond\_Prediction\_--\_Structuring\_Epistemic\_Integrity\_in\_Artificial\_Reasoning\_Systems

15. The Transformation of Epistemology: From Philosophy to ... \- Medium, https://medium.com/@mbonsign/the-transformation-of-epistemology-from-philosophy-to-mathematical-science-345969475c45

16. Epistemology in the Age of Large Language Models \- MDPI, https://www.mdpi.com/2673-9585/5/1/3

17. Semantic Laundering in AI Agent Architectures: Why Tool ... \- arXiv, https://arxiv.org/pdf/2601.08333

18. BEWA: A Bayesian Epistemology-Weighted Artificial Intelligence, https://arxiv.org/html/2506.16015v1

19. (PDF) A Neuro-Mathematical Blueprint for Epistemic Humility in Multi, https://www.researchgate.net/publication/395452892\_A\_Neuro-Mathematical\_Blueprint\_for\_Epistemic\_Humility\_in\_Multi-Agent\_AI

20. Towards Robot Scientists for autonomous scientific discovery \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC2813846/

21. A cut-free labelled sequent calculus for dynamic epistemic logic, https://academic.oup.com/logcom/article/30/1/321/5739324

22. Dynamic epistemic logic \- Wikipedia, https://en.wikipedia.org/wiki/Dynamic\_epistemic\_logic

23. Dynamic Epistemic Logic \- Stanford Encyclopedia of Philosophy, https://plato.stanford.edu/entries/dynamic-epistemic/

24. OPEN PROBLEMS IN LOGICAL DYNAMICS, https://archive.illc.uva.nl/lgc/papers/oqup.pdf

25. arXiv:2205.00784v2 \[cs.LO\] 12 May 2022, https://arxiv.org/pdf/2205.00784

26. On the Complexity of Dynamic Epistemic Logic ∗ \- arXiv, https://arxiv.org/pdf/1310.6406

27. Description logic \- Wikipedia, https://en.wikipedia.org/wiki/Description\_logic

28. OWL 2 Rules \- Foundations of Semantic Web Technologies, https://www.semantic-web-book.org/w/images/7/7a/OWL2-Rules-GeoS2009.pdf

29. Trust Revision for Conflicting Sources \- DTIC, https://apps.dtic.mil/sti/tr/pdf/AD1027413.pdf

30. Fusion of Subjective Opinions through Behavior Estimation, https://c4i.gmu.edu/\~pcosta/F15/data/fileserver/file/472251/filename/Paper\_1570107605.pdf

31. Dempster–Shafer theory \- Wikipedia, https://en.wikipedia.org/wiki/Dempster%E2%80%93Shafer\_theory

32. Delegated Reasoning and Epistemic Accountability in Human, https://www.preprints.org/manuscript/202601.1372

33. Integrating Internet of Things, Provenance, and Blockchain to, https://www.frontiersin.org/journals/sustainable-food-systems/articles/10.3389/fsufs.2020.563424/full

34. The Virtual Brain Ontology: A Digital Knowledge Framework for, https://www.biorxiv.org/content/10.1101/2025.11.19.689211v2.full-text

35. PROV-O: The PROV Ontology | Request PDF \- ResearchGate, https://www.researchgate.net/publication/320333535\_PROV-O\_The\_PROV\_Ontology

36. Structuring Epistemic Integrity in Artificial Reasoning Systems \- arXiv, https://arxiv.org/html/2506.17331v1

37. Semantic Provenance and Policy-Aware Explanations for Agentic, https://ceur-ws.org/Vol-4210/paper6.pdf

38. Reducing Aleatoric and Epistemic Uncertainty Through Multi-modal, https://www.researchgate.net/publication/404686903\_Reducing\_Aleatoric\_and\_Epistemic\_Uncertainty\_Through\_Multi-modal\_Data\_Acquisition

39. Complementing Self-Consistency with Cross-Model Disagreement, https://arxiv.org/html/2604.17112v1

40. (PDF) AI That Knows When It Is Wrong \- ResearchGate, https://www.researchgate.net/publication/412175285\_AI\_That\_Knows\_When\_It\_Is\_Wrong

41. The Art of Abstention: Selective Prediction and Error Regularization, https://www.researchgate.net/publication/353492014\_The\_Art\_of\_Abstention\_Selective\_Prediction\_and\_Error\_Regularization\_for\_Natural\_Language\_Processing

42. Selective prediction with calibrated uncertainty for retinal ... \- GitHub, https://github.com/ShahnawazKakarh/retinal-selective-prediction

43. Shapley Inconsistency Values \- AAAI, https://aaai.org/papers/kr06-027-shapley-inconsistency-values/

44. (PDF) Measuring Inconsistency through Minimal Inconsistent Sets, https://www.researchgate.net/publication/221393380\_Measuring\_Inconsistency\_through\_Minimal\_Inconsistent\_Sets

45. On the Measure of Conflicts: Shapley Inconsistency Values, http://www0.cs.ucl.ac.uk/staff/a.hunter/papers/hk10.pdf

46. Belief revision \- Wikipedia, https://en.wikipedia.org/wiki/Belief\_revision

47. (PDF) Belief Revision: an introduction \- ResearchGate, https://www.researchgate.net/publication/260134902\_Belief\_Revision\_an\_introduction

48. On the Logic of Theory Base Change: Reformulation of Belief Bases, https://ojs.aaai.org/index.php/AAAI/article/view/33636/35791

49. Logic of Belief Revision \- Stanford Encyclopedia of Philosophy, https://plato.stanford.edu/entries/logic-belief-revision/

50. Bayesian Epistemology with Weighted Authority \- arXiv, https://arxiv.org/pdf/2506.16015

51. Daily Papers \- Hugging Face, https://huggingface.co/papers?q=reliability%20decay

52. We Have No Satisfactory Social Epistemology of AI-Based Science, https://www.reddit.com/r/philosophy/comments/18um0tu/we\_have\_no\_satisfactory\_social\_epistemology\_of/

53. (PDF) Computational Models in Social Epistemology \- ResearchGate, https://www.researchgate.net/publication/317368847\_Computational\_Models\_in\_Social\_Epistemology

54. Multi-Source Fusion Operations in Subjective Logic \- arXiv, https://arxiv.org/pdf/1805.01388

55. Multi-source fusion in subjective logic | Request PDF \- ResearchGate, https://www.researchgate.net/publication/319370736\_Multi-source\_fusion\_in\_subjective\_logic

56. Subjective Logic Operators in Trust Assessment: an Empirical Study, https://www.researchgate.net/publication/259335092\_Subjective\_Logic\_Operators\_in\_Trust\_Assessment\_an\_Empirical\_Study

57. Subjective Logic Extensions for the Semantic Web \- Computer Science, https://www.cs.vu.nl/\~wanf/pubs/paper3.pdf

58. The Past and Future of AI Scientists | alphaXiv, https://www.alphaxiv.org/abs/2608.14407

59. The Robot Scientist Adam | Request PDF \- ResearchGate, https://www.researchgate.net/publication/224574580\_The\_Robot\_Scientist\_Adam

60. Parameterized Complexity Results for a Model of Theory of Mind, https://www.semanticscholar.org/paper/Parameterized-Complexity-Results-for-a-Model-of-of-Pol-Rooij/4b16adb4418cafe66d07902d932da83891a6df9b

61. On Verifying Expectations and Observations of Intelligent Agents, https://www.isichennai.res.in/\~sujata/papers/2022-ijcai.pdf

62. Climate Change Disinformation on Social Media: A Meta-Synthesis, https://www.mdpi.com/2076-0760/14/5/304

63. How well do models follow their constitutions? \- AI Alignment Forum, https://www.alignmentforum.org/posts/Tk4SF8qFdMrzGJGGw/how-well-do-models-follow-their-constitutions

64. Segment Anything with Robust Uncertainty-Accuracy Correlation, https://arxiv.org/html/2605.10603v1

65. Why Think for Yourself? | Episteme | Cambridge Core, https://www.cambridge.org/core/journals/episteme/article/why-think-for-yourself/64284D4FB6450ABCDF3D2F16AA0DD07F

66. Epistemic Environmentalism and Autonomy: The Case of, https://www.cambridge.org/core/journals/canadian-journal-of-philosophy/article/epistemic-environmentalism-and-autonomy-the-case-of-conceptual-engineering/36C72DA4101CFF4FDB5DA02E4637CB38

67. Toward a Perceptual Twins Protocol \- Jandislei Antonio Genova, https://jandisleigenova.com/artigos/testing-task-relative-epistemic-autonomy-perceptual-twins

68. Subjective logic \- Wikipedia, https://en.wikipedia.org/wiki/Subjective\_logic