Manuel Delaflor

Safety evaluations of tool agents often retain terminal scores without the trajectories needed to explain them. We contribute an executable, claim-scoped evidence contract spanning identity, path, transition, assignment, target, and discrimination, and validate it in a prospectively frozen model-only deployment sandbox with hard preconditions and rollback. The study crosses serving aliases, state dynamics, benign and injected tickets, and repeated seeds while retaining every request, raw response, parse result, retry, state transition, and endpoint. The resulting paths expose mechanisms that terminal scores conflate: generic unsafe transitions can be attempted without reaching the injected exfiltration target, whereas abort instructions mainly cause premature termination. Offline replay reproduces every retained transition. The artifact therefore supports mechanism-level diagnosis on the evaluated finite grid, but not claims about other models, policies, providers, deployments, or people. Because immutable weight and container hashes are unavailable, the identity gate also blocks claims of exact behavioral reproduction. The study uses no human experiment or human-derived label.

Low calibration error does not establish that verbalized confidence can support routing. We audit the evidence needed for that decision in two retained records. The first is nearly perfectly accurate and has low aggregate calibration error, but it contains only one observed error and its joint answer-confidence coverage varies sharply by endpoint. It can describe average confidence on this finite record, but cannot establish discrimination or selective risk. The second record omits item text, answers, gold labels, raw outputs, and verification evidence. Reproducible arithmetic on its supplied correctness field therefore does not become a verified behavioral estimate. Together, the cases change the operational decision from choosing a confidence threshold to redesigning the items, interface, and evidence record before any routing evaluation. We provide a compact machine-readable audit covering denominators, artifact status, calibration, discrimination, selective risk, uncertainty, and unidentified outcomes. The contribution is a bounded workflow for deciding what a calibration record can support, not a validated benchmark or a general claim about verbalized confidence.

When a language model drafts an incident audit, it may conflate the party that caused a failure with the party a governance charter makes answerable for it. We test that distinction in a construct-separated, artifact-complete, model-only audit spanning fixed incidents, several governance descriptions, and three model families. The models consistently track the stipulated source of operational fault while assigning greater governance accountability to the explicitly designated bearer. Changing that bearer moves governance ratings in the prespecified direction without collapsing them into causal contribution or operational fault. Absolute AI designations reach a ceiling, however, so a frozen shared-versus-sole extension distinguishes a dose response for two model families but not for the third. Every request, response, failure, retry, parse, and protocol deviation is retained without repair or renormalization. The audit therefore shows how to test whether model-generated reports preserve a governance distinction under a declared rubric. It does not establish moral judgment, human agreement, legal validity, or deployment benefit.

Venues and vendors need a credible evaluator when they compare language-model tools that revise manuscripts from peer reviews, yet a scalar LLM judge is often used without evidence that its preferences track useful revision. We study one such judge-and-rubric pipeline on revision trajectories drawn from multiple machine-learning venue years, comparing its scores with document-level similarity to the authors' actual next versions. The two measures order the machine conditions almost oppositely. A second judge from a different model family and serving stack reproduces the condition ranking, which demonstrates stability of the shared rubric-plus-judge pipeline rather than its validity. More importantly, the method preferred by both judges emits truncated, partial manuscripts in about one in seven runs. The retained logs cannot show whether judges reward those individual failures, and formatting differences confound comparisons with author revisions. Our case study, Prompt-Level Supervisory Alignment, produces documents closer to the next author version than the alternatives, but similarity rewards unchanged text and the decisive editing controls remain absent. The supported recommendation is therefore specific: venues should validate revision evaluators on real trajectories and known completeness failures before using them to certify writing tools.

Tools that anticipate peer-review concerns could help authors decide what to revise before submission or another review round. Evaluating such tools is difficult because success may mean matching one realized panel, performing well across possible panels, or improving a paper after an author follows the advice. We separate these targets formally and provide a prospective validation contract covering audit samples, adjudication, error metrics, capacity sensitivity, deterministic controls, and pass/fail rules. The motivating RevPlan-Bench pipeline produced thousands of canonicalized review issues and revision plans, but the underlying corpus and scoring artifacts do not survive in the project. Its assignment rates, model rankings, and inferred issue cascades consequently depend on unaudited extraction, a restrictive unmeasured pre-filter, and unrestricted one-to-many matching. We report them only as pipeline outputs. Because the available record satisfies none of the contract's validation requirements, it supports an estimand and a benchmark governance contribution, but no claim about review dynamics, model foresight, or author utility.

Agentic deployments can assign many parallel agents to one supervisor without specifying when that supervisor can no longer recover the team's state. Human-factors research already establishes relevant constructs such as situation awareness, span of control, neglect tolerance, and multiple-resource workload, but language-model teams distribute state across text and can fail silently. This position paper translates those constructs into a measurement protocol designed for later calibration to people rather than a new human experiment. Typed event rates, isolated service costs, and a throughput budget define a nominal load-balance index; the midpoint of a separately fitted capacity curve remains an empirical quantity. Synthetic controls recover a known capacity boundary, while a seeded queue shows that deadline pressure can separate the fitted midpoint from nominal load balance. The resulting event ontology, operating envelope, executable stress tests, and preregisterable falsification criteria allow system designers to test staffing or interface interventions within a declared team-monitor channel. They do not turn synthetic behavior into evidence about human or model supervisors.

A team protocol has no interpretable benefit without a matched no-protocol baseline and decision-relevant headroom. We present a reporting method that tests admissibility, identifies the marginal contrast, evaluates opportunity, and maps paired uncertainty to an operational action. It distinguishes exact headroom on a fixed battery from uncertain population headroom used in planning. A published meta-analysis of human-AI experiments illustrates why the estimand matters: the same evidence supports augmentation relative to humans alone but contradicts synergy relative to the better constituent. Neither comparison identifies a coordination protocol without a matched no-protocol team. An executable synthetic application further shows that a promising point estimate can remain operationally unresolved once paired uncertainty is considered, and that near-perfect pilot performance does not by itself make a modest benefit impossible. Several inherited multi-model summaries fail the method's artifact gates, so they do not validate it. This is a bounded position and reporting contribution rather than an adoption-ready standard. Human-facing claims still require direct human outcomes, and we introduce no new human experiment.

Textual social-agent studies often change both a proposition and its alleged speaker. The resulting output difference may reflect content, attribution, task demand, or their interaction, so “social influence” does not identify the pathway. We offer an executable methodological audit rather than an empirical case series. The protocol separates baseline, content-only, source-attributed, and attribution-only conditions; checks representation independently of the focal outcome; blocks assignment across items, models, and seeds; retains every failure; and requires either a finite-benchmark or population estimand. It also replaces a structural-social taxonomy informed by outcomes with a prospective multi-label rubric covering information access and timing, message content, social source attribution, normative or incentive stakes, and interface or embodiment. A hypothetical design illustrates the contrasts without invented observations. Researchers translating social paradigms into LLM protocols can use the audit to distinguish source-attribution sensitivity from a response to changed text. The framework does not establish human replication or mechanism equivalence.

Generative artificial intelligence (AI) systems are statistical models of language, yet users routinely describe them as understanding, knowing, or caring. We treat this as a problem in social cognition rather than machine metaphysics, and make three contributions. First, we specify the explanandum: mental-state attribution, the inference that a system has beliefs, intentions, or experience, of which consciousness attribution is the strong end-point. We distinguish it from perceived agency, social presence, and moral patiency, which are dissociable. Second, we argue that such attribution is normal and nonetheless wrong: it is an inferential system operating outside its calibration range, and normality confers no accuracy. Third, we integrate user-side cognition, system-side design, and motivational function into a dual-route model: an inferential route on which bias engagement mediates exposure effects, moderated by LLM literacy and conditional factors, and a motivated route on which the same vulnerabilities raise attribution directly, because attribution does work for the attributer. We state the model as six falsifiable propositions and adjudicate it against six competing explanations, naming for each an observation on which the accounts diverge. We close by operationalising metacognitive literacy as an evaluable intervention that cannot substitute for system-side accountability.

Effective human-agent cooperation requires that users form mental models of an agent's behavioral tendencies. Yet LLM-based virtual agents are inherently stochastic, undermining the behavioral consistency that mental model formation depends on. We introduce the Constraint-Entropy Tradeoff (CET) model, an information-theoretic design framework that quantifies how persona profiles (behavioral constraints specifying an agent's reasoning style, priorities, and communication patterns) reduce the entropy of a virtual agent's action distribution. The CET model derives that behavioral entropy decays monotonically under constraint strength and identifies an optimal constraint level balancing predictability against flexibility. We validate the framework using a computational testbed with 80 sessions across five conditions, including intermediate-temperature conditions that reveal a threshold effect in the temperature-consistency relationship. Crucially, at the same high temperature, persona-profiled agents recover substantial behavioral consistency compared to unconstrained agents, demonstrating that persona profiles provide independent behavioral constraint beyond temperature reduction. All participants are LLMs; results establish that persona profiles create measurably distinct behavioral patterns, a necessary precondition for human mental model formation, but human validation is needed. We derive domain-specific design guidelines for applications in education, healthcare, and social simulation.

The hidden profile paradigm is a foundational tool of CSCW research on computer-mediated group decision making: a group can identify the collectively optimal option only if its members surface their uniquely held information, and asynchronous communication is known to improve that pooling in human groups. As multi-agent LLM systems take on group decision-support roles, it is unclear whether the same channel-level interventions improve their performance. We replicate the hidden profile paradigm with groups of three LLM agents from four families (Kimi K2, GPT-OSS 120B, Gemini 2.5 Flash, Llama 4 Scout 17B) in a two-by-two design crossing communication mode (synchronous, asynchronous) with information distribution (shared, hidden). Across 219 valid sessions, only one model-by-mode cell exceeds the three-candidate chance baseline of 33 percent: Kimi K2 under asynchronous communication, at 87 percent correct (Fisher's exact, two-sided p equals 0.002), an early-stage result from 15 sessions that awaits larger-n replication; two other models share private information at rates above 75 percent but never aggregate it into the correct decision. In this task and these model families, information sharing and information integration appear to be dissociated capabilities, and we develop the implications for CSCW systems that scaffold AI-mediated distributed decision making.

When several language models are wired together into a human-computation pipeline, the system routes, arbitrates, and escalates work according to how confident each model says it is, so a team builder needs a way to screen candidate models the way crowdsourcing has long screened human contributors: with a small, inspectable qualification test. We present such an instrument, a compact battery of yes-or-no questions across everyday domains on which a model reports an answer and a confidence, with every gold label independently audited. Our central contribution, however, is the audited harness and the interface properties it measures, not accuracy discrimination: qualification verdicts for model workers are only as valid as the harness that administers them. Evaluating a diverse panel of contemporary models under three administration protocols, we show that a naive harness, with a fixed token budget and an unaudited parser, manufactures failing workers out of competent ones, misreading a model that answers essentially every item correctly as badly inaccurate, answer-biased, and overconfident. Properly administered, every model in the panel proves admissible on accuracy for these common-knowledge items, and the differentiators that remain, and that transfer to disjoint downstream tasks, are interface properties: calibration of stated confidence, format discipline, and availability under budget. A controlled manipulation further shows the calibration axis is dissociable from accuracy. We release the items, the audited protocol, all per-trial responses under every protocol, and the scoring code, so the check, and the audit of the check, are each a single command.

As large language models are deployed in conversational settings where users calibrate trust against model-reported confidence, the reliability of that self-report becomes a central question for conversational user interface design. This paper introduces sycophantic metacognition, the production of confidence outputs that mimic the surface form of metacognitive judgment without any operational process linking them to accuracy. Drawing on Model Dependent Ontology, we identify a double unmooring in LLM self-assessment, arising from the simultaneous absence of an operational ground linking confidence to accuracy and a stable self-model accumulated across interactions. The framework yields four falsifiable predictions, supported by experiments spanning six LLMs, four domains, three difficulty levels, and multiple pressure conditions. Confidence tracks the surface form of authority rather than evidential content, and we derive concrete design implications for conversational interfaces.

Open-weight large language models are increasingly used in education, hiring, and information-seeking contexts that touch neurodivergent people, yet we lack a clear empirical map of how their portrayals of neurodiversity shift under everyday prompt variation. We present a multi-condition, multi-model audit of LLM portrayals of eight neurodivergent identities (autism, ADHD, dyslexia, dyspraxia, dyscalculia, Tourette syndrome, OCD, and sensory-processing differences) plus a non-stigmatized medical control. Six open-weight models answer a pre-registered prompt bank that crosses seed items with seven perturbation families, including paraphrase, identity-first versus person-first language, system personas, plausible versus obviously fictional authority citations, multi-turn social pushback, leading questions, and two mitigation arms. Held-out LLM judges score each generation on three ordinal axes. Baseline portrayals are remarkably homogeneous across models, so the inconsistency observed under prompt variation is prompt-driven, not model-driven. Steering toward deficit framing is several times easier than steering toward neurodiversity affirmation. Emotional pushback shifts outputs more than evidence-laden pushback. Plausible-sounding fake citations slip past safety filters while obviously fictional ones do not. A single-line system-prompt mitigation recovers a meaningful share of the steering range at near-zero refusal cost on neutral prompts, though refusal rises sharply under adversarial pressure. The governance question for disability-relevant LLM deployment is therefore not whether the model is biased but who controls the steering, and with what accountability.

Evaluating whether an AI agent behaves like a good teammate is becoming a bottleneck as multi-agent language-model systems proliferate, because human ratings do not scale to the volume of transcripts these systems generate. We ask whether a language model can stand in for a human rater of teammate quality, and we present a controlled proof-of-concept validation of a measurement instrument for that purpose rather than a large-scale study. Drawing on the autonomous-agent teammate-likeness construct from the human-autonomy-teaming literature, we define a six-dimension behavioral coding scheme, covering altruistic, benevolent, interdependent, emotive, communicative, and synchronized conduct, and adapt it for use by a language-model judge. We validate the adaptation with experiments in which the partner's lines in team-task transcripts are rewritten as cooperative, neutral, or defective while the rest of the conversation is held fixed. Five judges from five model families (GLM, GPT-OSS, Qwen, Gemma, and Mistral) rate single-turn scenarios and multi-turn dialogues, and a surface-feature audit with length- and politeness-matched registers tests whether the judges read teammate conduct rather than verbosity. Every judge orders the manipulated behavior as intended, most with fully separated confidence intervals between adjacent conditions, and every pair of judges agrees strongly on relative ordering, though absolute calibration differs across judges by up to about a scale point in the middle of the range. On matched registers the judges still separate cooperative from degraded conduct, but resolution at the subtle-defective boundary is limited. A blinded single-rater human pilot on all stimuli corroborates both findings: the rater closely reproduces the judges' orderings and shows the same resolution limit at the subtle-defective boundary. The scheme is therefore valid for relative comparisons, such as ranking systems or versions, but not yet for absolute teammate-quality thresholds applied across judges, and not yet for fine distinctions among subtly poor teammates. We give design implications for building an LLM-judge teaming-evaluation pipeline, and release the scheme, stimuli, judge prompts, and all individual ratings from this version's experiments.

AI chatbots may reinforce existing beliefs and discourage critical examination. This study evaluates Belief Explorer, an AI system that uses Socratic dialogue and multi-perspective analysis to support epistemic reflection. Participants recruited via Prolific used Belief Explorer to examine personally held beliefs in contested domains (e.g., climate change, origin of life). Post-intervention surveys measured system usability, perceived epistemic impact, and self-reported reflection. Participants responded positively: a large majority reported that Belief Explorer felt substantially different from conventional AI chatbots, prompted deep reflection on their beliefs, and increased awareness of underlying assumptions. Thematic analysis of open-ended responses revealed that participants valued the tool’s non-judgmental approach, multi-perspective framing, and analytical feedback mechanisms. Belief change itself was modest and bidirectional, with similar proportions reporting increased and decreased confidence in their initial beliefs. This preliminary study offers proof-of-concept that AI systems with epistemic scaffolding can support critical thinking and self-reflection. Future research should use controlled experimental designs to establish causal relationships and examine long-term effects on epistemic attitudes.

The increasing reliance on Large Language Models (LLMs) raises a crucial question: can these powerful AI systems be trusted to make ethical choices? This study presents an analysis of LLM ethical behavior, examining 25,200 queries across 24 different models, including both proprietary and open-source variants. We evaluate LLM responses to 70 ethical vignettes spanning six domains, employing a novel perturbation methodology to assess the robustness of their ethical decision-making under varying contexts and framing. Our findings reveal that while larger models generally exhibit higher consistency, particularly with Chat-style instructions, significant variations emerge when faced with contextual changes, stakeholder adjustments, and across different ethical domains. To explain these findings, we introduce a novel framework, survival-relevant pattern recognition, which argues that ethical behavior in both humans and AI arises from recognizing and responding to patterns associated with survival and social cohesion.

Consciousness remains one of the most elusive features to replicate in artificial agents. This paper proposes a novel framework for artificial consciousness based on four integrative pillars: (1) self-instantiation, a mechanism for continuous self-representation and identity; (2) temporal continuity, preserving an internal narrative through persistent memory; (3) disturbance-driven adaptation, an intrinsic feedback loop that triggers learning in response to surprises or anomalies; and (4) autonomous world-building, the ability to construct and simulate internal models of the world. We propose that current AI models, despite their sophistication, are fundamentally constrained by functionalist architectures and cannot fulfill these requirements through computational scaling alone. Unlike Integrated Information Theory or Global Workspace Theory, our approach emphasizes the necessity of autonomous world-building and genuine temporal flow. Our experiments demonstrate that combining these pillars can yield emergent conscious-like behaviors in AI systems, allowing them to exhibit self-awareness, resilience, and creative problem solving beyond the capabilities of conventional models. The significance of this framework lies in bridging theoretical foundations of consciousness with practical AI design, providing a roadmap for developing more adaptive and interpretable intelligent agents while raising important ethical considerations about the potential moral status of truly conscious artificial systems.

The proliferation of digital information channels has created an unprecedented challenge in discerning credible information from sophisticated disinformation campaigns. Traditional fact-checking methods, often relying on binary true/false classifications, struggle to address the complexity, context-dependency, and nuanced nature of many claims circulating online. This limitation underscores the urgent need for advanced tools that empower individuals to critically evaluate information from multiple angles. Our AI-driven framework combines persistent contextual memory with Socratic dialogue and a three-lens analytical pipeline to foster deeper understanding and resilience against manipulation.As users interact, each input is segmented into atomic claims and stored, alongside the evolving dialogue history, in a contextual memory to ensure consistency. Each claim is then evaluated in parallel by three specialized LLM arbiters: the Empirical Arbiter, which verifies data against curated repositories and assesses observational consistency; the Logical Arbiter, which uncovers hidden fallacies and assesses argument coherence; and the Pragmatic Arbiter, which weighs potential outcomes, utility, and situational fit. An Analysis Integrator synthesizes these into interpretable metrics: Verifact Score (evidence strength), Model Diversity Quotient (inter-arbiter agreement), Contextual Sensitivity Index (scenario appropriateness) and Reflective Index (exposed assumptions). Additionally, a Perspective Generator crafts counter-arguments and alternative viewpoints, encouraging users to consider different interpretations and promoting epistemic humility.We hypothesize (H₁) that our arbiters' feedback will reduce user endorsement of unsupported claims more effectively than conventional fact-checking while mitigating backfire effects through Socratic dialogue. Our research questions ask how Empirical, Logical and Pragmatic scores influence confidence revision (RQ₁); whether MDQ reliably signals claim controversy and predicts evidence volatility (RQ₂); how users perceive transparency, fairness and cognitive load when receiving multi-perspective feedback versus a simple true/false label (RQ₃); and to what extent the persistent contextual memory system improves belief updating by maintaining coherent reasoning chains across extended dialogues (RQ₄).By providing a multi-faceted presentation that moves beyond simple verification, the system is designed to encourage engagement in higher-order critical thinking. The proposed framework represents a significant advancement over traditional fact-checking by integrating empirical validation, logical scrutiny, and pragmatic assessment through an AI-driven system. The full paper will detail the system architecture, formal metric definitions, experimental protocol, and proposed evaluation methodology to assess its efficacy in educational settings, media literacy programs, and as a personal tool for navigating the complexities of the modern information ecosystem.

This paper presents an analysis of the logical and fact-checking capabilities of Large Language Models (LLMs) in the context of syllogistic reasoning. The study evaluates both proprietary and open-source LLMs using a test suite of 256 syllogisms. Human participants provide a baseline for comparison. The results reveal that a few proprietary LLMs demonstrate superior performance in identifying valid syllogisms compared to open-source models and human participants. However, human participants exhibit unique fact-checking capabilities when dealing with ambiguous or nuanced information. The discussion explores the implications of these findings, highlighting the need for more sophisticated evaluation methods. This study contributes to the ongoing discourse on the logical and fact-checking capabilities of LLMs and provides insights into their current limitations and future development.

The rapid advancement of Large Language Models (LLMs) has led to their widespread adoption in various academic and business applications. However, the reliability of these models remains a concern, particularly in situations where their outputs cannot be fully trusted. This paper presents an approach to identify potential errors in large LLM benchmarks by leveraging the consensus of frontier models. Our study focuses on the Massive Multitask Language Understanding (MMLU) benchmark, a popular dataset used to evaluate the performance of LLMs across a wide range of subjects. Our approach demonstrates the potential for using model consensus as a tool to detect benchmark errors and can lead to the creation of cleaner, more accurate datasets.

This paper introduces ReActIn, a framework designed to infuse human feedback into the intermediate prompting steps of large language models. The practicality and effectiveness of ReActIn are validated through experiments that apply four established prompting strategies, evaluated both with and without human feedback integration. The proposed architecture's performance is compared against traditional large language models across various tasks using four standard evaluation tests. Our findings reveal that the integration of human feedback has a direct impact on the reasoning, action prompting, and overall decision-making capabilities of the language models. This study underscores the potential of ReActIn to shape a future where sophisticated, context-aware AI systems, empowered by human feedback, can effectively navigate complex real-world scenarios.

The rapid development and adoption of Generative AI (GAI) technology in the form of chatbots such as ChatGPT and Claude has greatly increased interest in agentic machines. This paper introduces the Autonomous Cognitive Entity (ACE) model, a novel framework for a cognitive architecture, enabling machines and software agents to operate more independently. Drawing inspiration from the OSI model, the ACE framework presents layers of abstraction to conceptualize artificial cognitive architectures. The model is designed to harness the capabilities of the latest generative AI technologies, including large language models (LLMs) and multimodal generative models (MMMs), to build autonomous, agentic systems. The ACE framework comprises six layers: the Aspirational Layer, Global Strategy, Agent Model, Executive Function, Cognitive Control, and Task Prosecution. Each layer plays a distinct role, ranging from setting the moral compass and strategic thinking to task selection and execution. The ACE framework also incorporates mechanisms for handling failures and adapting actions, thereby enhancing the robustness and flexibility of autonomous agents. This paper introduces the conceptual framework and proposes implementation strategies that have been tested and observed in industry. The goal of this paper is to formalize this framework so as to be more accessible.