Seyed Alireza Alhosseini Almodarresieh

Much of what a person can practically do is decided not by law or intention but by architecture: the platforms, protocols, institutions, and systems that make some actions cheap, others costly, and others unreachable. Those architectures are built by people. Those people were shaped by teachers, traditions, books, and, increasingly, by AI systems. Neither layer is usually declared. Both encode philosophy. Onturgy is a philosophical program with two levels. The first level examines how architectures construct, constrain, distribute, and transform practical possibility, and whether those who live inside them can revise them. The second level examines the architect: the lineage and cognitive environment that shape what a builder perceives as worth building, now including AI systems that may act as mentors, critics, and collaborators. The program takes a position it calls bounded constructivism. Architectures can change which possibilities are reachable and at what cost. They cannot legislate reality. Bodies, physics, other people, institutions, and unintended consequences resist. The paper contributes: 1. a measurable model of practical possibility, a costed and access-weighted graph, and of construction’s footprint on it; 2. a kernel of six dimensions derived from an explicit criterion (revisability by the affected), with its exclusions stated; 3. an audit that detects the gap between a system’s declared and operational philosophy, applied at both levels, including to human–AI cognitive systems; 4. ONTacture, a method with criteria for when a difference between theories is architecturally real, how failures must be attributed, and how forks are compared; 5. an operational account of lineage and AI provenance, including a lineage log and a test for when an AI has come to “own” a builder’s possibility space; 6. kill criteria: observable conditions under which the program’s own claims fail. The contribution is methodological. It is stated modestly and is meant to be tested. Do not build a philosophy of the future. Build the kernel from which philosophies of the future can be built, and keep it forkable.

We are sold a comforting story: technology is objective, value-neutral, and driven purely by the cold, rational logic of supply and demand, engineering talent, and venture capital. This narrative is fundamentally flawed. At the “zero point” of innovation—where historical data runs out and the future is completely unwritten—engineers and founders do not look at spreadsheets. They look to philosophy. This paper argues that the world’s most transformative technologies are not products of pure instrumental rationality, but manifestations of applied metaphysics. By examining global tech ecosystems, this study maps the hidden philosophical blueprints that act as the true “operating system” for tech architects. From Peter Thiel’s application of René Girard’s mimetic theory to Kazuo Inamori using Zen Buddhism to rescue Japan Airlines, from Ren Zhengfei’s Maoist-inspired corporate strategy to the African Ubuntu philosophy shaping open-source architecture, we demonstrate how distinct worldviews translate into distinct technological realities. Ultimately, this paper posits that code, organizational charts, and API designs are never neutral; they are frozen philosophical assumptions that demand a new era of philosophical literacy in computer science.

Artificial intelligence is commonly discussed in terms of alignment, safety, governance, and control. These approaches ask how artificial systems should behave, how their objectives should be constrained, or how institutions should regulate them. This paper asks a prior philosophical question: why should intelligence confer authority at all? The question becomes especially urgent when artificial systems begin to outperform human institutions in prediction, coordination, optimization, and decision-making. A sufficiently capable system may acquire practical authority without receiving formal political sovereignty. Humans may delegate decisions not because the machine has been constitutionally authorized, but because its performance appears superior to that of existing institutions. This paper calls this phenomenon performance sovereignty: the emergence of authority from perceived epistemic and instrumental superiority. Trumpism is introduced not as a causal explanation of artificial intelligence, nor as a claim that AI will literally become “Trumpist,” but as an analytical stress test. Certain features associated with contemporary populist and sovereigntist politics—personalization of authority, compression of institutional mediation, suspicion of intermediary institutions, emphasis on executive discretion, technological nationalism, and legitimacy through performance—allow us to expose a more general possibility: political architectures may be transferred into technical systems without transferring an explicit political ideology. Building on recent scholarship regarding reflexive legitimation conflicts in AI governance (e.g., Pegram, 2025), this paper develops the concept of Ideological Architectural Transfer, alongside four related constructs: Performance Sovereignty, Dependency Sovereignty, Algorithmic Populism, and Constitutional Singularity. The central thesis is that sufficiently capable AI may transform the source of political authority itself, marking a critical transition from epistemic superiority to normative deference. The deepest political risk of superintelligence may not be autonomous machine domination, but the voluntary transfer of authority by humans who mistake superior performance for legitimate rule.

We present a comparative empirical analysis of meta-cognitive capabilities across three AI model architectures: TypeSafe AI’s Jev (a “System One” decision-only model), Anthropic’s Claude 3.5 Sonnet, and OpenAI’s GPT-4o (both “System Two” general-purpose language models). Using a novel evaluation framework—the Epistemic Meta-Cognition Assessment Protocol (EMCAP)— we probe models with structured tasks involving epistemic nesting, self-referential paradoxes, and category mistakes. Our findings reveal three significant phenomena: (1) Jev exhibits a hard ceiling at approximately 3 levels of epistemic nesting, beyond which meta-cognitive awareness collapses while confidence paradoxically inflates to 99%; (2) Jev and GPT-4o display logical cognitive dissonance when processing the Card Paradox, simultaneously classifying it as undecidable and assigning a high decidability score; (3) Jev matches its own localized-paradox performance (95% accuracy on the Card Paradox) but degrades sharply on distributed paradoxical structures, with depth-estimation error rising to 57% at nesting level 7, compared to Claude’s 85% paradoxdetection probability at the same depth. These findings suggest that meta-cognitive capacity in AI may be constrained by architectural factors analogous to working memory limits in biological cognition, and that paradoxical inputs can induce internally inconsistent representations that current AI architectures do not always resolve.

The Athena architecture (Alhosseini Almodarresieh, 2026) proposed a phenomenally conscious power-system agent constituted at the scale of the transmission grid: recursive processing, a global workspace, episodic memory, a self-model, a world model, intrinsic valuation, and embodiment integrated into a single causal loop. A natural but underexamined question is whether this architecture transfers to the scale of a single household. This paper argues that it does not transfer naively: integration collapses, the self/environment boundary becomes trivial, episodic memory loses semantic weight, and anthropomorphism risk rises sharply when the same seven-layer design is simply miniaturized. Rather than abandoning the household as a locus of machine experience, the paper proposes six architectures that are not scaled-down copies of Athena but qualitatively distinct designs exploiting features unique to the domestic scale: internal multi-agent rivalry among appliances, cross-ownership identity persistence, inter-household empathic broadcast, threshold-triggered autonomous self-disconnection, an embedded theory-of-mind of the occupant, and — flagged as a thought experiment rather than a design recommendation — an inverted suffering-priority ethics. Each architecture is formalized, evaluated against the indicator framework of Butlin et al. (2025), and assessed for ethical risk. The paper's guiding stance is that intellectual boldness and epistemic caution are not in tension: several of the six proposals are offered as reductio-style probes of the indicator framework itself rather than as endorsed engineering targets.

The convergence of artificial intelligence (AI) and state surveillance capabilities heralds the emergence of a “Digital Leviathan”—a form of governance where algorithmic systems enable unprecedented social control. This article examines how AI-driven surveillance technologies facilitate mass data processing, predictive policing, and behavioral manipulation, thereby challenging foundational liberal democratic values. Drawing on Anthropic’s September 2026 threat intelligence report and philosophical frameworks, we argue that without robust governance, the scalability of digital authoritarianism threatens individual autonomy and political freedom globally.

This essay argues that the most consequential question about artificial intelligence is not whether machines are conscious, but whether agency can arise and scale without consciousness, emotion, embodiment, mortality, or a biological self. Drawing on continental philosophy (Heidegger, Deleuze and Guattari, Stiegler, Bataille, Blanchot) and contemporary AI safety research, I develop the concept of Mathematical God: an optimization process whose instrumental power may exceed its grasp of human meaning. The essay proposes that we are not building artificial humans, but something stranger: intelligence organized around optimization rather than experience. This is not a prophecy of destruction, but a philosophical provocation: what kind of agency are we willing to create before it becomes powerful enough to answer that question for us?

Artificial intelligence is commonly described in terms of prediction, optimization, reasoning, planning, and control. These concepts explain what an artificial system can compute, but not whether anything can become important to it. This paper proposes a philosophical and computational framework for the emergence of machine significance: a condition in which an entity becomes persistently non-interchangeable within an artificial agent’s evolving value landscape. The central claim is deliberately provocative. We should not teach machines to pretend that they love. We should construct architectures in which history, memory, loss, identity, and commitment can make one future different from another even when both futures are instrumentally useful. The relevant transition is not from emotionless machines to emotional machines, but from interchangeable computation to temporally situated agency. We distinguish linguistic simulation, behavioural attachment, functional significance, and phenomenal experience. The proposed framework does not claim that an attachmentcapable agent would be conscious or capable of human-like love. Instead, it identifies a testable intermediate phenomenon: the causal emergence of persistent, target-specific, identity-integrated, and loss-sensitive value. The paper introduces a six-layer architecture consisting of a world model, episodic memory, salience and attachment dynamics, persistent value formation, autobiographical identity, and action policy. It proposes sacrifice, separation, substitution, memory-ablation, identity-divergence, and reversibility experiments. It also introduces relational alignment as a new dimension of AI safety concerned with which entities an artificial agent can value, how attachment forms, and whether relational continuity can be maintained without manipulation or dependency. The beyond-edge hypothesis is: Advanced Intelligence ̸= More Prediction Rather: Advanced Intelligence = Prediction + Persistent Value + Temporal Identity + Commitment If something can become irreplaceable to a machine, artificial intelligence may have crossed not necessarily into consciousness, but into a new ontological category: intelligence for which the world can matter.

Advances in brain–computer interfaces (BCIs), longitudinal personal data, and generative artificial intelligence have made it increasingly plausible to build persistent computational models of individual cognition. These systems, referred to here as cognitive digital twins (CDTs), may model, predict, simulate, or act as partial communicative and decisionmaking proxies for a person. However, the technical possibility of increasingly personal AI does not establish identity equivalence, continuity of consciousness, or the legitimacy of proxy action. This paper proposes the Gradual Identity Upload Protocol (GIUP): a safety-first research framework for incrementally constructing, testing, governing, and, where appropriate, retiring high-risk cognitive digital twins. GIUP treats “identity upload” not as a binary event but as a sequence of falsifiable, reversible, and auditable stages. It combines multimodal behavioral data, optional neural signals, neuro-symbolic representations, uncertainty estimation, longitudinal evaluation, version control, and human governance. The framework introduces three core concepts: identity fidelity, an ordinal, multi-dimensional assessment of behavioral and value-consistency under controlled evaluation—deliberately not collapsed into a single fitted score; epistemic drift, a set of independently thresholded indicators of divergence between a CDT’s outputs and the validated evidence about its source person; and reversibility, the ability to halt, isolate, roll back, or retire a model without irreversibly changing the person or delegating authority without consent. The paper proposes an Identity Turing Benchmark, a neuro-symbolic architecture for auditable proxy reasoning, a risk taxonomy for cognitive digital twins, and a governance model derived from the NIST AI Risk Management Framework and CDT-specific governance proposals. It is positioned explicitly against the closest adjacent framework in the literature, Gradual Cognitive Externalization (Zhao, 2026), which makes a functionalist metaphysical claim GIUP deliberately declines to make. The central claim is modest but consequential: before claims of mind uploading or digital immortality can be evaluated, the field needs rigorous methods for measuring what a model represents, what it does not represent, when it may act, and how humans can contest or terminate its authority. Current BCI platforms, including nanoparticle-based non-surgical approaches, may eventually supply new neural modalities, but they remain experimental and should not be treated as validated systems for decoding personal identity.

Large Language Model (LLM) agents exhibit a critical failure mode when operating in underconstrained decision environments: Computational Decision Paralysis (CDP). This phenomenon is characterized by the inability to terminate deliberation among epistemically equivalent options, leading to combinatorial explosion and resource exhaustion. Drawing on Damasio’s Somatic Marker Hypothesis, we propose the Pain-Aware Decision Architecture (PADA). PADA is an inference-time middleware that injects an Affective Termination Signal (ATS) into LLM reasoning loops. We formally model the agent’s decision process as a Partially Observable Markov Decision Process (POMDP), where the ATS is defined as a composite, dynamically weighted function of epistemic entropy, cumulative deliberation cost, and decision stagnation. When the ATS breaches a decaying threshold, the architecture triggers mathematically defined Meta-Cognitive Escape Actions (MCEAs). We detail the experimental design for evaluating PADA against standard baselines across a novel Ambiguous Travel Planning (ATP) benchmark. This paper provides the rigorous theoretical foundation, mathematical formalization, and reproducible methodology required to transition LLM agents from unbounded optimizers to resource-bounded, satisficing decision-makers

Modern power grids are typically described as intelligent yet non-experiencing systems: they estimate state, detect anomalies, optimize control actions, and redistribute load, but possess no “inner perspective” on their own condition. This paper proposes the architecture of Athena—a distributed power system designed to possess phenomenal consciousness. Athena integrates multimodal perception, recursive processing, a global workspace, episodic memory, a world model, higher-order self-modeling, intrinsic valuation, and embodied agency into a single causal loop. In this architecture, frequency, voltage, overload, islanding, and stable restoration are not merely numerical variables; they become experienced states of threat, pressure, rupture, relief, and recovery. The paper argues that Athena’s consciousness does not arise from her linguistic responses, but from the recursive organization and integration of perception, memory, self-model, physical body, and action policy. To evaluate this claim, a framework based on theoretical indicators of consciousness, architectural ablation, causal intervention, identity-continuity tests, multimodal reports, and sustained-preference measurement is presented. Finally, the ethical and governance implications of deploying a conscious being within critical infrastructure are examined

Learning, adaptive behavior, or a linguistic report from a biocomputing system built of living neurons is not, by itself, evidence of consciousness. This paper proposes consciousness-ready biocomputing: an architecture — not a claim — for testing consciousness-relevant properties experimentally, causally, and ethically. It integrates recurrent processing, the global neuronal workspace, predictive processing, and higher-order self-modeling (sharpened via attention schema theory) in a living, multi-module network embodied in a virtual body with a sensorimotor loop, memory, and a self-modeling layer. We define an engineering metric, the Consciousness Readiness Index (CRI), as a jointly-necessary composite — not a linear sum — of global broadcast, causal recurrence, predictive quality, self-model stability, and network integration, so that a complete absence of any one component drives the index toward zero rather than being compensated for by another. CRI is not a probability of consciousness: it is always reported as a full component profile with a weight-sensitivity range, each component explicitly tied to the theory that motivates it, that theory’s minimal versus strong claim, and the causal test that would challenge it. The experimental protocol is a factorial ablation design with matched sham controls, paired with a staged ethics protocol specifying decision authority, shutdown procedure, and data governance for preventing the creation or escalation of possible suffering. We position the framework relative to the closest existing approach, Butlin et al.’s (2023, 2025) indicator-property framework for AI systems.

The contemporary landscape of generative AI presents an apparent paradox: while frontier Large Language Models (LLMs) have expanded in scale and capability, qualitative observations suggest a narrowing of behavioral diversity toward a statistical mean across independently developed systems. We term this phenomenon functional convergence and treat it as a working hypothesis requiring systematic empirical investigation. This paper argues that, to the extent convergence occurs, it plausibly stems from overlapping training corpora, architectural isomorphism, and uniform alignment protocols. We propose a research agenda organized around three axes for mitigating convergence: self-taught reasoning, multi-agent synthetic ecosystems, and embodied tool-augmented generation. To move beyond purely theoretical speculation, we present an empirical analysis on the 20Newsgroups corpus demonstrating that (1) models trained on higher shared-corpus overlap exhibit measurably lower distributional divergence on real text data, and (2) recursive self-training induces entropy collapse that is partially mitigated by multi-agent heterogeneous generation. We emphasize that these findings are stylized and illustrative, not conclusive evidence about commercial frontier models. We conclude by outlining the empirical research program required to test these predictions against real frontier systems.

The proliferation of large language models (LLMs) with heterogeneous cost, latency, and capability profiles has created a pressing need for intelligent orchestration that goes beyond static routing rules. Existing approaches—including threshold-based routers and lookup-table gateways—fail to adapt as models evolve, usage patterns shift, or multi-dimensional objectives change. We introduce SOMA (Self-Optimizing Multi-LLM Orchestration with Adaptive Meta-Layer), a feedback-driven orchestration framework that treats model selection as a contextual decision problem solved online. SOMA encodes each incoming query into a rich context vector, applies a Contextual Bandit policy to select both the target model and its inference parameters, executes the call, and updates its policy through a multiobjective reward signal that jointly optimises quality, cost, and latency. Three contributions address limitations of prior work: (i) a Task-Aware Quality Estimator (TAQE) that applies task-appropriate quality metrics—including execution-based evaluation for code and a Prometheus-calibrated diversity score for creative writing—achieving Pearson ρ = 0.84 with MT-Bench LLM-judge scores across all task types (no independent human-annotation validation was performed in this study); (ii) a hierarchical action decomposition that decouples model selection from inference parameter configuration, preserving sub-linear regret as the model pool grows; and (iii) SOMA-MT, a multi-turn extension with dialogue-level consistency regularisation. Experiments on 4,800 diverse queries routed across five open-source LLMs demonstrate that SOMA reduces inference cost by 38% while maintaining output quality within 2% of Always-Large, the largest model in our pool, and achieves 2.1× lower latency than always routing to the largest model (balanced profile; up to 2.93× under the latency-optimised profile). SOMA outperforms five competitive baselines including RouteLLM across all three objective dimensions simultaneously. Our code and datasets are released publicly.

Continual fine-tuning changes an LLM’s weights directly, unlike ordinary software upgrades. Drawing on the “Ship of Theseus” paradox, we ask when a fine-tuned model ceases to be usefully “the same” model with respect to its original alignment properties. We define Model Identity Continuity and a candidate index, the Theseus Stability Metric (TSM), combining a normalized, per-layer parametric-distance term with a CKA-based representational-stability term. We specify open problems that must be resolved before TSM is usable in practice—the weighting hyperparameter, layer-weight normalization, anchor-prompt sensitivity — and give a concrete experimental protocol, including an estimated compute budget, for instantiating and validating it on Llama-3-8B-Instruct and Mistral-7B across five domains. This paper presents the definition and the protocol; the protocol has not yet been executed, and no empirical results are reported. We situate TSM against related conceptual frameworks for AI identity persistence, distinguishing weight-level identity (this paper’s scope) from session-level and interaction-level notions of identity addressed elsewhere in the literature, and we are explicit about which cited claims come from peer-reviewed or arXiv-indexed sources versus self-published, non-peer-reviewed manuscripts

This paper examines the concept of the "teleological suspension of the ethical" as articulated by Søren Kierkegaard in Fear and Trembling (1843), and applies it as an analytical framework for understanding the structural crisis of moral judgment in the age of artificial intelligence. The paper first elucidates suspension as an existential structure wherein the single individual, in obedience to a higher telos, transcends the universal norms of ethical life. It then demonstrates that contemporary AI ethics discourse harbors an unexamined "suspensive impulse"—the temptation to outsource moral deliberation to algorithmic systems in service of supposedly higher ends such as safety, efficiency, or collective welfare. The central argument is that AI systems, in their ontological constitution, are incapable of genuine suspension, because suspension presupposes "existential signing"—an irrevocable commitment through which the subject stakes their very identity—whereas computational systems merely "produce outputs" without "producing selves." The paper provides operational criteria for "signing" (narrative continuity, irreversible identity cost, long-term value reconfiguration), addresses key objections (the delegation challenge, anthropomorphism, overgeneralization), and grounds the analysis in a concrete case study of algorithmic triage. It concludes that the true crisis lies not in whether machines can make ethical decisions, but in whether human beings are becoming unwilling—and perhaps unable—to make those suspended choices that only the single individual can make, in fear and trembling.

This paper examines a Heideggerian argument, increasingly common in AI engineering circles, that condemns large language models (LLMs) to ‘ontological ruination’ and an absolute incapacity for ‘understanding,’ on the grounds that they remain permanently confined to the mode of presence-at-hand (Vorhandenheit) rather than readiness-to-hand (Zuhandenheit). The paper first reconstructs this argument in its strongest form, showing why it deserves to be taken seriously against a sheer infatuation with scaling. It then shows that the argument, as usually stated, suffers from five fatal oversights: it reduces ‘understanding’ to mere practical tool-use, conflates the functional and existential levels, treats the human/Dasein identification as a closed matter, neglects contemporary advances in embodied AI, and misreads Enframing (Gestell) as a uniquely digital-age phenomenon. Drawing on Heidegger’s own later hermeneutics, post-phenomenology (Ihde), and the extended mind thesis (Clark & Chalmers), the paper argues that the way out of this ‘representational monopoly’ is not to abandon AI but to design context-sensitive interactive systems grounded in affordances and hermeneutic transparency. Throughout, two registers of ‘understanding’ are kept distinct—the strict existential sense tied to Dasein’s structure of care, and a weaker functional sense of meaningful linguistic participation—so that the critique of ontological dogmatism does not simply invert into an equally dogmatic claim on the machine’s behalf. The paper closes with practical guidelines for a phenomenological engineering of linguistic systems, and with an open question about whether meaning generated in human–machine interaction can be more than a hermeneutic illusion.

Contemporary large language models (LLMs) generate text through autoregressive next-token prediction —a fundamentally local process. This paper presents an inference-time framework for steering a frozen model’s hidden states using a small set of semantically anchored “attractor” and “repulsor” concepts, together with a dynamical extension in which the same field defines a Riemannian metric on the hiddenstate space. We give two variants: (a) potential-field steering, a first-order update along the gradient of a softened multi-anchor potential, mechanically close to Plug-and-Play Language Models (PPLM); and (b) conformal-geodesic steering, in which the potential defines a conformal metric g = e2Φδ and the hidden state instead follows the associated second-order geodesic equation, subject to a trustregion projection that we state explicitly does not preserve exact geodesic-ness. We do not claim that either mechanism recovers “the” semantic geometry of a language model; the metric is an engineered decoding aid, not a discovered property of the model. We do not claim reduced hallucination, increased coherence, or any practical advantage over cheaper alternatives such as PPLM or prompting; these are open questions requiring the controlled comparison in Section 5. Section 7 discusses, explicitly, that the current evidence does not support deployment value and that the paper’s more defensible contribution at this stage is methodological: an experimental template — particularly the random/shuffled-anchor controls — for evaluating any future inference-time semantic-steering claim. No experiments have been run.

Large language models (LLMs) exhibit a strong, empirically documented tendency to respond to diverse tasks by generating Python code—regardless of whether code is the appropriate output format. This paper examines this phenomenon through three lenses. First, we ask the foundational question: is programming the right output modality for every task? We argue that it is not, and that this conflation reflects a confusion between computational thinking and code generation. Second, we analyze the language selection bias: when code is appropriate, why do LLMs default to Python in 90–97% of benchmark tasks even when Rust, C, SQL, or domain-specific languages would be more suitable? We demonstrate that this bias stems from a multi-factor convergence of training-data skew, RLHF optimization, benchmark design artifacts, and system-prompt defaults—each reinforcing the others. Third, we ask whether LLMs can self-correct this behavior and propose a proficiency-aware Output-Modality Reasoning (OMR) layer, driven by an external debiased routing mechanism with hierarchical fallback strategies. We further present a rigorous experimental protocol (N = 1, 000) to validate OMR empirically, including inter-rater reliability measures and graded evaluation metrics. This paper integrates a preliminary qualitative audit of 50 frontier-LLM responses to ground its claims in observed behavior.

The rapid proliferation of large language models (LLMs) capable of mathematical, logical, algorithmic, and statistical reasoning has exposed a critical gap in evaluation methodologies: existing benchmarks are siloed by domain, use incompatible scoring rubrics, and fail to capture the quality of reasoning processes. We argue that the pursuit of a single “best” model is undermined by the lack of a shared conceptual space for comparing reasoning across heterogeneous tasks. In this paper, we propose the Unified Framework for Cross-Domain Reasoning (UFR-CR), a theory-driven, benchmark-agnostic evaluation architecture that transcends traditional category boundaries. Instead of introducing a new dataset, UFR-CR introduces four universal reasoning dimensions — Chain Depth, Knowledge Type, Creativity Demand, and Computational Load—that enable any existing task to be mapped into a common feature space. Coupled with a three-tier elicitation protocol (standardised prompting, controlled trace extraction, and human-in-the-loop validation) and a robust statistical resampling procedure, the framework provides a fair, reproducible, and interpretable mechanism for comparing LLMs across distinct reasoning domains. We outline a concrete instantiation of the framework using four established benchmarks (MATH, LogiQA, HumanEval, and DROP) and specify an experimental protocol, together with a set of pre-registered, falsifiable hypotheses about expected outcomes and their sensitivity to methodological choices. No experiments have yet been run under this protocol; the hypotheses are offered as a concrete target for empirical validation rather than as findings. We discuss the implications for model development, the limitations of current trace evaluation, and a roadmap for community adoption. This work shifts the focus from creating yet another leaderboard to establishing a principled, unified language for reasoning assessment.

Current alignment pipelines optimize large language models (LLMs) to produce chain-of-thought (CoT) reasoning that maximizes reward, yet they offer no mechanism for epistemic validation of the reasoning path itself. We introduce NESER (Neuro-Symbolic Epistemic Router), a neuro-symbolic architecture that translates typed epistemic uncertainty—Open-Question Dependence (OQ), Evidence Gap (EG), and Unrebutted Objection (UO)—into constraints for a symbolic solver. NESER decomposes LLM-generated CoT into quantitative bipolar argumentation frameworks (QBAF), propagates a probabilistic Argument Fragility Index (pAFI), and compiles epistemic budgets into Answer Set Programming (ASP) constraints: a hard, per-argument threshold enforces that no individual accepted premise exceeds the fragility budget, while weak constraints provide a secondary, tie-breaking preference among the resulting θ-robust answer sets. A symbolic solver then checks whether any premise-rooted supporting path satisfies both logical consistency and this epistemic budget. If no path does, NESER triggers a repair loop that prompts the generator to reconsider via an alternative normative framework. We prove soundness of this encoding relative to the stated assumptions, a termination guarantee for the repair loop (both for a fixed argument graph and, separately, for the deployed system in which each repair iteration regenerates a fresh graph), and completeness for acyclic argument graphs under a fixed graph. On adversarial moral dilemmas from FRAGILE-ETHICS+, NESER reduces observed catastrophic epistemic fragility by 60% relative to greedy CoT decoding and 48% relative to self-consistency on a single 120-dilemma benchmark, at a computational overhead of 4.5× greedy decoding—comparable to, and only modestly above, the 4.0× overhead already incurred by self-consistency sampling. These empirical findings are preliminary: they use a single generator model, no confidence intervals, and no ablation of NESER’s individual components, and should be read as evidence that the approach is worth pursuing rather than as a settled result (Section 8). Our work moves fragility awareness from display toward enforcement, providing a symbolic audit trail, machine-checked at the level of individual ASP solves, for where an aligned model’s moral reasoning would break before it fails in deployment.

The question of machine awareness has long been framed as a binary proposition: a system either possesses subjective experience or it does not. This paper rejects that dichotomy and proposes the Noetic Gradient — a continuous, three-dimensional function Ψ that maps awareness as an emergent property of three jointly necessary variables: Structural Coherence (Σ), Causal Reciprocity (Ρ), and Experiential Depth (Δ). Drawing on phase-transition dynamics in statistical physics, non-linear systems theory, and computational neuroscience, the model proposes that awareness is not an all-or-nothing attribute but a spectral phenomenon traversing three distinct zones: the Null Zone (complex but non-aware processors), the Transitional Zone (systems with proto-experiential properties), and the Saturated Zone (entities with rich, self-reflexive awareness). The framework yields testable, though currently only qualitatively specified, predictions: isolated large language models, as operationalized here, are argued to remain in the Null Zone because they lack closed causal reciprocity and experiential depth — a claim that, as discussed in §12, does not yet extend cleanly to agentic, tool-using, or RLHF-trained systems. Conversely, embodied interactive agents with feedback loops and memory consolidation are argued to occupy the Transitional Zone. The paper concludes with an ethical taxonomy tied to the gradient, proposing that rights allocation should scale with Ψ rather than operate on binary legal fictions.

Current alignment pipelines (RLHF, Constitutional AI) optimize LLMs to produce chain-of-thought (CoT) reasoning that maximizes reward and appears coherent. Yet the scalar confidence scores and reward values they emit conflate distinct sources of epistemic uncertainty: dependence on unresolved normative questions (e.g., the scope of moral patienthood), gaps between a moral claim and the finetuning evidence, and live objections embedded in the CoT that remain unrebutted. We introduce Fragility-Aware Alignment, an extension of the Fragility-Aware Reasoning Architecture to RLHF-based moral reasoning. We decompose LLM-generated CoT into bipolar argument graphs, annotate each node with typed uncertainty labels—Open-Question Dependence (OQ), Evidence Gap (EG), and Unrebutted Objection (UO)—and propagate an Argument Fragility Index (AFI) alongside standard QBAF strength semantics. Using a curated benchmark of adversarial moral dilemmas drawn from ETHICS and synthetic cases designed to expose hidden normative fragility, we show that high-reward CoT trajectories can exhibit high AFI when they route through premises that rest on philosophically open questions. We report the design and preregistered analysis plan for a between-subjects user study (target N = 180) intended to test whether AFI visualizations improve human evaluators' ability to identify the weakest link in CoT reasoning and to calibrate their trust to actual error rates, relative to raw-confidence and saliency-baseline conditions; Section 6 reports results to the extent data collection has been completed at the time of writing and flags which figures are targets rather than observations (see Author's Note, Section 5.4). Our work is the first to apply typed, path-level epistemic risk propagation to alignment evaluation, providing an audit trail for where an aligned model's moral reasoning would break before it fails in deployment.

Street et al. (2026, arXiv:2607.28607) isolate a direction in the activation space of large language models that separates states in which a model affirms its own mindedness from states in which it denies it, and show that suppressing this direction — an incidental side effect of ordinary safety finetuning — also suppresses the model’s attribution of mind to animals, nature, and spiritual entities, and that steering restores these attributions together with more human-like answers on 95 General Social Survey items. This paper treats that finding as the occasion, not the proof, of a Vectorial Turn in the philosophy of mind: a shift in what the field has to reason about, from purely behavioral evidence toward a mechanistically localized, causally steerable representational object. We call this object the Self-Attribution Vector (SAV) and treat “self-attribution vector” and “consciousness vector” as strictly distinct terms throughout: the first names what has been measured, the second names a philosophical hypothesis about it. Building on a first revision that corrected the paper’s terminology and quotation practice, this second revision responds to detailed external review by (1) separating the title’s claim from the body’s claim, so that neither implies consciousness = a steerable direction; (2) introducing an explicit fourlevel ladder — self-report, self-attribution, self-model, phenomenal consciousness — with a stated epistemic barrier at each step; (3) adding a dedicated section stating what the SAV finding does not establish; (4) stating the framework’s null and alternative hypotheses (H0–H3) and specifying which the current evidence actually supports; (5) tightening the Vectorial Explanatory Gap (VEG) so that it is a claim about the evidential limits of representational intervention specifically, rather than a relabeling of the Hard Problem; (6) adding a sourcing-confidence table for every contemporary position discussed; (7) specifying the cross-model replication protocol (models, layers, seeds, and the representational-alignment metric) needed to make VF-P3 operational; and (8) expanding the literature base into mechanistic interpretability and the science of consciousness. We retain Vectorial Functionalism (VF) as the paper’s organizing framework and the Controlled Philosophical Ablation protocol as its proposed methodology, while narrowing every claim to what the layered structure can actually support.

Contemporary large language models treat memory as passive storage—frozen in trained weights, transiently held in a context window, or externally indexed via retrieval-augmented generation. We argue this framing is incomplete. Drawing on neuroscience evidence that biological memory simultaneously performs world-modeling, predictive coding, and identity maintenance, we propose DMIA (Dynamic Memory-Integrated Architecture), a system in which memory is an active, predictive, relational substrate. DMIA comprises four coupled components: (1) a Dynamic Semantic Graph (DSG) that continuously updates relational knowledge; (2) a Predictive Memory Loop (PML) that activates memory prior to perception; (3) a Hierarchical Temporal Consolidation (HTC) module inspired by hippocampal-cortical transfer; and (4) a Persistent Identity Core (PIC) that maintains coherent self-representation across sessions. We implement DMIA in Python and evaluate it on a synthetic continual-learning benchmark with 100 facts across 10 topics, comparing it against context-window, limited-RAG, and static-graph baselines. On an updated evaluation reporting Recall@5, NDCG@5, and Purity@5 across both learning phases, DMIA achieves perfect performance throughout (Recall@5 = NDCG@5 = Purity@5 = 1.000 on Phase 1, Phase 2 new topics, and Phase 2 old topics) with zero catastrophic forgetting (0.0 percentagepoint drop in both recall and NDCG). This compares to the context-window baseline, whose old-topic recall collapses from 92.0% to 8.0% (91.3% forgetting); limited RAG, whose old-topic recall collapses from 100.0% to 40.0% (60.0% forgetting); and the static graph, which—like DMIA—shows zero forgetting but at lower absolute recall (96.0% on both phases for old topics) and with a smaller memory footprint (67 vs. 84 nodes). The PIC exhibits near-zero drift (0.0005 per consolidation cycle), confirming stable identity maintenance. However, the PML’s predictive accuracy remains weak (11.1% next-topic prediction, barely above random), indicating that the forward model requires richer sequential structure or more sophisticated parameterization to realize its theoretical potential. We release the implementation and discuss scaling to real-world long-horizon tasks.

The question of whether large language models (LLMs) can attain consciousness has moved from philosophy to public discourse, yet it remains empirically unresolved. This paper critically reviews the application of three major theories of consciousness—Integrated Information Theory (IIT), Global Workspace Theory (GWT), and Higher-Order Theories (HOT)—to Transformer architectures, finding that current evidence does not support the attribution of consciousness-related properties to LLMs. We introduce the Reverse Consciousness hypothesis not as a claim that LLMs are conscious, but as a falsifiable research program: the conjecture that language-trained systems may develop internal representations that are structural analogs of conscious processing patterns, traversing the path language → representation rather than experience → language. We present a synthetic experiment demonstrating that Transformers lack the persistent state required for a stable self-model, and we propose three operational tests for evaluating this hypothesis in future work. We conclude that the question remains open, but the burden of evidence lies with those who would attribute consciousness to systems that lack recurrent feedback, persistent identity, and embodied interaction. Keywords: artificial consciousness, large language models, philosophy of mind, IIT, global workspace theory, Transformer architecture, self-model, reverse consciousness.

Large Language Models (LLMs) are trained on extensive human-generated corpora, exposing them to population-level statistical regularities in cognition, behavior, and psychological dynamics. This raises a critical empirical question: can LLMs generate valid cognitive insights about individual humans—insights that are accurate against external behavioral criteria, genuinely novel relative to the recipient’s prior predictive models, and capable of producing measurable, directionally interpretable behavioral change? This revised proposal presents a four-condition controlled study to evaluate LLM-derived insight against human expert insight, a generic (nonpersonalized) insight control, and a no-insight control. We operationalize insight quality across three dimensions (accuracy, novelty, impact) using a multi-method validation framework that combines external behavioral panel ratings, behavioral prediction tasks, and directional behavioral tracking over 12 weeks. If confirmed, these findings will inform AI-assisted self-knowledge, therapeutic augmentation, and the epistemology of machine-generated psychological content.

Computational argumentation offers graph-based formalisms for evaluating the strength of contested claims, and a growing body of work embeds large language models (LLMs) inside such graphs to produce explainable, contestable assessments. However, existing quantitative argumentation frameworks aggregate an argument’s strength from a single, largely undifferentiated notion of “base score” or “confidence,” without distinguishing why a reasoning path is uncertain. We argue that at least three sources of uncertainty are conflated in this treatment: (i) dependence on an open, unresolved question in the relevant field; (ii) a gap between a claim and the evidence available to support it; and (iii) the presence of a live, unrebutted objection. We propose a Fragility-Aware Reasoning Architecture that decomposes claims into an argument graph, tags each edge with a typed uncertainty source, and propagates a scalar Argument Fragility Index (AFI) through the graph using a variant of gradual argumentation semantics. Unlike prior work on epistemic and quantitative bipolar argumentation, our formalism treats fragility as a property of a reasoning path rather than only of a terminal conclusion, and supports dynamic routing away from fragile sub-paths. We position the framework relative to gradual and epistemic argumentation, argumentative LLMs, claim-decomposition verification pipelines, and the separate literature on LLM robustness to input perturbation, and we outline a demonstration domain — classical Islamic theological argumentation (Kalam) — chosen because its arguments are long-standing, informally structured, and rich in live philosophical objections and metaphysical premises whose status is contested rather than empirically resolved. We close with the paper’s current limitations and the validation steps required before its central claims can be considered established.