Agentic artificial intelligence systems — autonomous, multi-step AI agents capable of planning, tool use, and cascading real-world action — present a qualitatively distinct governance challenge from static AI models. The EU AI Act, while a landmark regulatory achievement, contains a structural gap: it mandates documentation and incident reporting but does not require real-time, publicly verifiable, tamper-proof audit infrastructure adequate for governing agentic systems at the pace and scale of current deployment. This paper documents a pattern of AI-enabled harm across four independent evidential sources — the ENISA 2025 Threat Landscape report, the November 2025 GTG-1002 autonomous cyberattack campaign, the February 2026 breaches of Mexican democratic infrastructure, and concurrent AI-automated attacks at scale — and argues that this pattern establishes the governance case for mandatory real-time accountability infrastructure for critical agentic systems. We propose a three-pillar framework. First, a Public Immutable Audit Ledger (PIAL): a distributed ledger-anchored system recording cryptographically hashed event logs in real time, governed by a technology-neutral requirements framework specifying fourteen functional and non-functional criteria any qualifying platform must satisfy. Second, a revised incident taxonomy separating automated telemetry — immediate, machine-generated — from narrative disclosure obligations, resolving the perverse incentives created by conflating these in existing frameworks. Third, a tiered implementation pathway classifying agentic systems into four risk tiers (Critical, High-Risk, Standard, Experimental) using an operational decision framework, with obligations scaled proportionately. The paper identifies zero-knowledge proof capability as a domain-specific precondition — not merely a research priority — for Tier A PIAL adoption in healthcare and law enforcement contexts where existing legal obligations under GDPR Article 9 and Directive 2016/680 may not be satisfied by current architecture. Six specific legal questions requiring formal resolution by the EU AI Office are identified, spanning GDPR Chapter V data transfers, NIS2 Article 23 interaction, DORA Article 19 alignment, and the data sovereignty status of public distributed ledger anchor submissions. The framework is accompanied by a reference implementation case study and a companion Technical Blueprint. The governance infrastructure proposed is proportionate, deployable with existing technology across the core architecture, and designed to be compatible with the EU AI Act's existing provisions while addressing their identified limitations.
[Depreciated and replaced by V3] The application-specific clean rebuild has not yet been published; its authoritative theoretical boundary is now the governing V3 branch: After Turing: The Fold Machine - An Exact, Parameter-Free and Machine-Closed Derivation of Classical Computational Science from Smithian Fold Theory; From Fold to Consciousness: An Exact, Zero-Parameter and Machine-Closed Foundational Reconstruction of Consciousness and Cognitive Science from Smithian Fold Theory. The V3 source platform is https://github.com/MettaMazza/ernos-labs-sft-platform. The original DOI, concept DOI, version number and files are preserved for transparent historical provenance; this record must not be presented or cited as current V3 work. v4.0 — the word-scale gap closes within the fold. Rung 5e (pre-registered): the fold-factor mixing law — every context level that holds contributes, weighted 2^level, the engine's own forced halving constant — carries the pure counted engine, with no twin, no prose flood, zero training and zero parameters, past the gradient-trained transformer at word scale: cross-entropy 3.1907 vs the same-day twin's 3.4292 (replicated across two independent anchorings; stacked with the Rung 5d extraction: 3.1344). Both scales of the task gate now belong to the counted engine. Rung 5d's transfer-in verdict is SUPPORTED across three independent arena anchorings in one day. New in the architecture: tool graduation (acts held, values never — a question territory that a tool answered once runs the tool itself thereafter, fresh), recall as regeneration across every memory tier, and judge-independent graduation scoring. End-to-end verification: 36/36. v3.4: Rung 5d, the transfer-in — pre-registered verdict SUPPORTED: the trained twin's dyadically-loud fold content is extracted and installed INTO the counted engine as a counted prior with zero new parameters, closing 55.6/87.9/101.4% of the available gap at k=16/32/64 while the random-truncated null closes 10.1/24.5/56.5%; at half budget the loud shape beats the full twin's own. The word-scale rematch is recorded in full (twin retrained on today's text; decomposition included). Also: judge-independent graduation scoring (boot-discovered pool, cycle-parity alternation), multi-orbit binding (XI-4 in full), recall-is-regeneration (a held experience re-walks its own orbit, never reprinted), the public SOTA table beside the local giants with cited published figures, and one-command replication kits (GPT-2 weights auto-fetch; 13/13, 39/39 proven on a fresh clone). End-to-end verification: 36/36. Full paper v1.1 — supersedes the pre-paper (From One Axiom to Master-Level Chess — and the Law Inside Neural Networks). Built from scratch by one woman, working alone, in under twenty-four accumulated hours: where a score falls short it marks an implementation gap at measurement time, never a limit of the mathematics — the gains between releases are the finding. v1.4 adds the fold eye (vision as exact integer Walsh spectra, self-certified by integer Parseval per image, recognition of seen images with no image model in the loop) and the graduation score (blind head-to-head vs the teacher, tallied per question-territory; the teacher retires as wins cross the majority lock) -- and documents the 2026 convergence: DeepSeek Engram arrives at deterministically-addressed exact memory from the gradient side, and two independent results place the optimal curriculum at p = 1/2, the fold lock. v1.6: the full omnimodal engine (the voice via Kokoro, the fold ear -- sound as Parseval-certified integer Walsh spectra, video composed from frames + sound), speaker-transparent reasoning threads, and 32/32 end-to-end empirical verification of the entire architecture including persistence across process death. v1.7: removal-proof omnimodality, measured -- every supporting model is a teacher with an exit: a sound taught once by the synthesis teacher is re-spoken from the engine's own exact counted record in 0.00s with no model; a sound heard once is recognized natively with no transcriber; 34/34 end-to-end verification. v1.9: zero-model perceptual learning (the human observer -- a novel image learned and re-recognized at share 1.00 with no model in the loop); agentic self-knowledge (the observer reads the engine's own source, measured); the hourly progress instrument with a committed pre-boot birth line; one-tap y/n closure. v2.0 (flight-ready): the full modern-agent toolkit (live web search/fetch, paginated reading, in-file grep -- every call held as a training trace), the 43-domain everything-curriculum under the fold-only law, SOTA 1-1 benching on the public MMLU test split with the newborn baseline committed, generation closure (the Learning Law reaches generate() itself), and 36/36 end-to-end verification. v2.1: the ReAct law (reason-act-observe enforced in-turn; narrated intent without an act is detected and forced), reasoning trained on the observer's NATIVE thinking tokens (STaR-gated) with both minds' full thinking streamed to the user, and document intake (a sent file is reading -- inboxed, counted, persistent). v2.2: the identity stated correctly -- UnisonAI is an OMNI MODEL (language, sight, hearing, speech, and video on one held memory), not a language model; LLMs remain the contrast class only. v3.0: the full-altitude rewrite -- the complete omni model documented at the same depth as the spectral science: thirteen sections, the architecture organ by organ with every measurement, Rung 5c as its own section, the empirical record and its committed birth line, 36/36 end-to-end verification, and the 2026 convergence. This paper is a PROOF of The Smithian Fold Theory of Everything, not the main event: the theory (one axiom, zero free parameters, 1,844 machine-verified forced checks) is at DOI 10.5281/zenodo.21182469 and github.com/MettaMazza/Smithian-Fold-Theory-Of-Everything -- run the prover yourself. The engine: github.com/MettaMazza/UnisonAI. v3.3: the LLM-native presence suite -- the exact registered protocol applied to GPT-2's entire knowledge-storage class: 13/13 tensors, 39/39 checks, unanimous (margins 3.4-79.3x); the flagship claim now rests on the flagship objects, with diffusion/speech models recast as cross-domain breadth. Three connected results and the architecture they force. First, a pre-registered, self-certifying spectral instrument shows trained neural-network weights carry placement-law in the dyadic (Walsh) basis: 18/18 unanimous on validated released models; the law concentrated in transformer expansion projections and token embeddings across three unrelated architectures (up to 230x chance in GPT-2), attention at chance; strictly training-caused (He-initialised controls at 1.0x); surviving 4-bit deployment quantization. A recipe map from 124M to one trillion parameters shows the law tracks training recipe, not scale or architecture — strongest carrier DeepSeek-R1-671B at 43–47x — and loud-recipe weights transform under the fold's transformation group exactly as solved game-theoretic value fields do. Second, the "learned similarity space" is a counted object: word kinship as exact co-occurrence shares reproduces semantic family structure (quark → lepton, neutrino, proton) with zero parameters and zero gradients. Third, UnisonAI: a complete language architecture in which every LLM mechanism — memory, attention, similarity, learning, prediction, generation — is replaced by a machine-verified law of the Smithian Fold Theory, zero trained parameters end to end. On identical held-out text the fold-native engine outperformed its trained transformer twin (cross-entropy 1.289 vs 1.888) after reading the corpus once (26 seconds) against 48,000 gradient readings (21 minutes per seed). Deployed as a live, continuously-learning agent whose teaching loop also runs autonomously: a teacher model asks, judges, and closes the learning law itself, and the engine self-plays against its own held lessons. Negative results reported in full with their scopes. Companion to The Smithian Fold Theory of Everything (DOI: 10.5281/zenodo.21182469; 307 suites, 1,844 forced checks, 0 failures). Engine and records: github.com/MettaMazza/UnisonAI and github.com/MettaMazza/Smithian-Fold-Theory-Of-Everything.
Open access
8 source records
Explainable Artificial Intelligence (XAI)
Generative Adversarial Networks and Image Synthesis
An Explainable Audit Trail (EAT) records process execution traces of an agentic workflow. EAT can enable an organisation to efficiently remain compliant with regulations by presenting its audit results to it. However, current state of the art in EAT lacks privacy protection of agent's models which can be intellectual properties of the organisation. Exposing audit trails to external entities may facilitate orchestration of attacks on the agent's model. Auditing a complex workflow will require verification of dependencies among tasks as preconditions to execute a task. Further, it is necessary for the audit algorithm to ensure that a process execution trace follows the pre-planned process execution model for security reasons, i.e., audit should include functionality that can check if the agents have deviated from its planned process execution models. In this paper, we build a secure EAT that can address these gaps in the state of the art in EAT for agentic workflows. Our main contribution is the application of zero-knowledge-proof on verifying audit procedures. It proves the audit has validated correctness of chain-of-thoughts, the execution trace at the runtime matches the planned process execution , and complete traceability among logs of a complex workflow involving dependencies among the tasks in terms of preconditions. Our solution provides a trust-less infrastructure to verify the audit results to external entities while not exposing the audit trails. We used lattice-based zero knowledge proof for this procedure. We provide an analysis on the EAT procedure. We show experimental evaluation of the EAT with workflow dataset.
This appendix proposes the Luevano Standard as an assurance architecture combining zero-knowledge proofs for model inference with remote attestation, providing stronger runtime evidence for AI governance. It frames this as a technological measure supporting EU AI Act Article 9 and Article 11 compliance demonstration, not as a universal legal solution.
Machine learning-based intrusion detection systems can identify malicious network activity, but their predictions and explanations are typically accepted without verifying that they were derived from the same input. This thesis develops a public-model/private-input zero-knowledge framework for certifying a prediction and its semantic explanation while keeping the processed network-flow features private. The framework is instantiated through a TON_IoT intrusion detection case study in which 104 processed features are mapped into five semantic groups. Logistic Regression is used as the proof-compatible public model, while XGBoost provides a stronger plaintext performance baseline. The main technical contribution is an implementation-backed proof relation that jointly verifies Logistic Regression inference and an ordered top-3 semantic explanation from the same private input. Under a fixed training-mean reference, semantic-group Exact SHAP for the linear score reduces to a direct group-wise weighted sum, enabling its implementation in a Circom circuit and verification using Groth16. The quantized relation achieves more than 99.99% prediction agreement with the floating-point model, while ordered top-3 explanation agreement is approximately 93.8%. Valid proofs are accepted, whereas incorrect predictions, malformed rankings, and out-of-range inputs are rejected. The results demonstrate the feasibility of cryptographically binding a prediction and a semantic explanation under private tabular inputs. The implemented relation remains limited to a public linear model, fixed semantic groups, and an approved reference vector, and does not provide arbitrary-model explanation verification, model confidentiality, or production-ready provenance.
Non-fungible tokens (NFTs) have become a key asset class in Web3 markets, where visual artwork, textual narratives, and on-chain transaction patterns jointly determine value, yet their pricing dynamics remain volatile, opaque, and difficult to explain. Existing NFT valuation methods typically either ignore the multimodal nature of NFTs or treat assets as independent samples, failing to exploit the rich relational structures induced by shared creators, collections, and ownership patterns, and offering limited interpretability for high-stakes financial decisions. To address these challenges, we propose NFT-Insight, a multimodal graph transformer framework that unifies visual, textual, and blockchain information on a heterogeneous NFT graph and explicitly links structural and content signals to valuation behavior. The framework identifies closely related NFTs via a joint similarity measure in the multimodal embedding space, propagates information through a relation-specific graph attention network and a global transformer encoder, and adopts a regularization strategy that encourages consistent valuations for highly similar assets while still allowing data-driven differentiation. In addition, NFT-Insight integrates attention-based and SHAP-based explanations into a unified analysis pipeline, enabling joint study of valuation behavior and feature attributions at the level of related NFT pairs. Experiments on three large-scale, real-world NFT datasets show that NFT-Insight consistently outperforms strong unimodal, multimodal, and graph-based baselines, reducing MAE and RMSE by up to about 20% in static valuation (withR2up to 0.904), achieving robust cross-market performance with averageR2≈ 0.84 andr≈ 0.93, and attainingR2= 0.911 in temporal forecasting. Interpretability analysis reveals that visual, textual, blockchain, and graph-relational features achieve a high alignment between SHAP importance and attention weights (average Spearman correlation above 0.8), and case studies highlight meaningful valuation patterns driven by rarity, speculative trading, and temporal market shocks. Overall, the proposed framework offers a multimodal graph–based perspective on explainable NFT valuation and market forecasting, and provides a general template for incorporating complex relational and content interactions into graph-based learning in decentralized digital economies.
While Decentralized Autonomous Organizations (DAOs) and Artificial Intelligence are reshaping the governance of academic societies, reliably integrating on-chain decisions with off-chain physical activities remains a critical challenge. The fundamental bottleneck is the difficulty of reliably integrating real-world execution outcomes into the digital decision-making loop. To address this, we propose an endogenous contribution evaluation framework integrating Decentralized Physical Infrastructure Networks (DePIN) and Vision-Language-Action (VLA) models. This approach maps physical entities to on-chain decentralized identities. By leveraging VLA edge nodes to analyze multimodal behavioral data collected via DePIN, the system autonomously generates a verifiable Proof of Real-World Contribution (PoRWC). This proof subsequently drives on-chain incentive distribution through a reputation-weighted consensus mechanism. Consequently, this framework establishes an endogenously trustworthy closed loop from physical processes to digital governance. We demonstrate its feasibility and scalability through a case study of the Chinese Association of Automation (CAA), providing a robust engineering path for the parallel governance of modern academic societies.
Decision-OS V5 Revised (SiriusA2) addresses a practical AI safety problem: how human oversight can prevent irreversible decisions from being executed under pressure, confusion, coercion, or panic. It is designed for safety-critical, non-medical decision support settings where a user may be authenticated, yet the execution path may still be unsafe. The framework proposes a human-in-the-loop confirmation layer for irreversible risk. Instead of allowing a valid user action to move directly into execution, SiriusA2 routes protected actions through auditable confirmation states such as Request, Observe, Hold, Approve, Reject, Stop Candidate, Execute, and Revoke. The core mechanism is a trajectory-aware duress_score. This score is not an intent classifier, diagnosis, truthfulness score, or autonomous veto. It is an operational control-routing signal used to detect when a valid execution path deviates from an ordinary decision trajectory and approaches irreversible harm. SiriusA2 preserves human final consent through two-step confirmation, an explicit revoke path, optional family multisig, and a Zero-Knowledge approval layer (310/320) that verifies authorization without exposing personally identifiable information. The duress_score does not replace ZK approval or multisig; it routes actions into the confirmation path, while ZK qualification and multisig provide independent authorization conditions before irreversible execution. The revised manuscript integrates the SiriusA Adoption Gate into the main paper. The gate provides a Hold-first confirmation path for irreversible, externally pressured, unusually urgent, or high-stake actions: Request → Observe/Hold → Approve/Reject → Execute/Revoke. A score-based Stop Candidate does not automatically become Stop or Freeze. Stop or Freeze requires an independent non-score condition, such as verified revoke input, ZK-qualified m-of-k approval, policy-defined guardian confirmation, or an emergency protocol condition. If no such condition is available, SiriusA intentionally prefers continued Hold and evidence preservation over an AI-only execution veto. This release also clarifies cold-start behavior, causal bridge support, baseline maturity, corrected event terms, disclosure boundaries for calibration-sensitive parameters, and non-PII audit requirements. Safety is operationalized through auditable state transitions, non-PII KPIs, five-line gate outputs, evidence packaging (ZIP + SHA256), and explicit prohibitions on automatic transmission, payment, or reporting. A minimal proof-of-concept gate exists as a public runtime artifact, demonstrating PASS / DELAY / BLOCK routing, conservative severity merging, evidence union, pre-execution checking, and auditable JSON outputs. Deployment-level validation, calibration, and domain-specific robustness remain future work. Gateway / series index:https://github.com/shin4141/decision-os-paper Recommended read order:V5 Revised (SiriusA2) → V6 (PIC) → V8 (v2)Optional: V7 (AGI definition) Related repositories:- V5 Policy Pack / specification and adoption materials: https://github.com/shin4141/paper-public- Gate engine / MMAR-L0: https://github.com/shin4141/mmar-l0-core- SiriusA core runtime: https://github.com/shin4141/siriusA-core SSOT:GitHub repository “decision-os-paper”.This PDF corresponds to the revised SiriusA2 release candidate committed to the SSOT repository. Release note:This revised release integrates the trajectory-aware duress_score definition, SiriusA Adoption Gate, ZK qualification layer, family multisig, independent non-score Stop conditions, non-PII audit structure, V4-compatible escalation interface, and proof-of-concept gate positioning into the main paper. Transparency / Author’s Note:https://github.com/shin4141/decision-os-paper/blob/main/AUTHORS_NOTE.md
The immutable nature of smart contracts necessitates rigorous auditing, especially for ERC compliance, to prevent significant economic losses. While automated tools, particularly those combining Large Language Models (LLMs) with symbolic execution, have improved detection, they often suffer from false positives, false negatives, and insufficient interpretability. This paper introduces SymExplainer, a novel integrated framework designed to overcome these limitations. SymExplainer features an LLM-Enhanced Rule Semantic Extraction Module that deeply understands ERC specifications and misuse patterns using multi-stage prompting and a domain-specific knowledge base. Its Context-Aware Symbolic Execution Engine then efficiently prioritizes exploration paths based on these LLM insights. Crucially, a Violation Verification and Interpretability Generation Module performs secondary LLM-based cross-validation to significantly reduce false positives and produces comprehensive, natural language reports detailing "why," "where," and "how-to-fix" confirmed violations. Evaluated on a ground-truth dataset of 159 expert-annotated ERC violations, SymExplainer achieved perfect recall with zero false negatives and substantially reduced false positives to only 15, outperforming state-of-the-art methods like SymGPT (which reported 29 false positives and 1 false negative). An ablation study confirmed the critical contribution of each module, and qualitative human evaluation validated the high clarity, accuracy, and actionability of its interpretability reports. Despite a modest increase in computational cost, SymExplainer provides a more precise, reliable, and transparent solution for smart contract auditing through unparalleled accuracy, reduced noise, and actionable insights.
Understanding the economic intent of Ethereum transactions is critical for user safety, yet current tools expose only raw on-chain data or surface-level intent, leading to widespread "blind signing" (approving transactions without understanding them). Through interviews with 16 Web3 users, we find that effective explanations should be structured, risk-aware, and grounded at the token-flow level. Motivated by these findings, we formulate TxSum, a new user-centered NLP task for Ethereum transaction understanding, and construct a dataset of 187 complex Ethereum transactions annotated with transaction-level summaries and token flow-level semantic labels. We further introduce MATEX, a grounded multi-agent framework for high-stakes transaction explanation. It selectively retrieves external knowledge under uncertainty and audits explanations against raw traces to improve token-flow-level factual consistency. MATEX achieves the strongest overall explanation quality, especially on micro-level factuality and intent quality. It improves user comprehension on complex transactions from 52.9% to 76.5% over the strongest baseline and raises malicious-transaction rejection from 36.0% to 88.0%, while maintaining a low false-rejection rate on benign transactions.
As fraud patterns in the Ethereum ecosystem become increasingly sophisticated, traditional detection methods face limited generalization capability and insufficient interpretability. Although Large Language Models (LLMs) possess powerful semantic understanding and reasoning abilities, their direct application in fraud detection still suffers from critical issues, including inadequate domain knowledge integration and scarcity of high-quality interpretable training data. To address these challenges, this paper proposes Large Language Model for Transaction Anomaly Detection (LLM-TAD), a framework that constructs interpretable training data through a dual interpretation strategy combining XGBoost with SHAP/LIME to provide complementary feature-level insights, and achieves domain knowledge injection and capability optimization via a two-stage approach involving supervised fine-tuning and instruction fine-tuning. Experimental results demonstrate that the proposed method achieves a fraud detection accuracy of 93.01% and an explanation quality (BERTScore) of 0.7939, achieving synergistic improvement in both accuracy and interpretability.
Prescriptive analytics seeks to identify optimal interventions for achieving desired outcomes, with causal inference playing a pivotal role in assessing intervention impacts on complex systems. However, existing approaches frequently neglect critical data privacy considerations and provide no means to verify the integrity of their recommendations. These limitations hinder its adoption in high-stakes domains such as healthcare and finance. In this paper, we introduce, zkCLEAR, a zero-knowledge proof (ZKP)-based C ausal Inference ( LEA rning and R easoning) framework for privacy-preserving and verifiable prescriptive analytics. Our solution allows data owners or service providers to cryptographically prove the validity of prescriptive conclusions derived from causal analysis without disclosing sensitive source data or proprietary causal models. We develop a suite of ZKP-friendly causal operators to build efficient causal modules, including structure learning, parameter learning, probabilistic inference, and counterfactual reasoning. To optimize performance, we also introduce a workflow decomposition strategy to facilitate efficient proof generation for complex workloads. We demonstrate the utility of zkCLEAR through three real-world applications. The framework faithfully follows the behavior of non-ZKP counterparts, with moderate overheads for privacy and verifiability. Additionally, we evaluate its efficiency and scalability using real-world datasets. It shows up to a 35.1× speedup in proof generation time and a 214.5× reduction in proof size compared to current general-purpose ZKP systems.
This paper addresses the critical need for accountability in artificial intelligence (AI) systems, particularly in domains where decisions have significant societal and ethical implications. We propose a novel framework leveraging auditable attestations to ensure provable compliance with predefined standards and regulations. The core of our approach involves generating verifiable proofs about the behavior and characteristics of machine learning models, allowing for independent audits and assessments. We explore the theoretical foundations of such attestations, focusing on cryptographic techniques like zero-knowledge proofs and secure multi-party computation, which enable the verification of model properties without revealing sensitive information. Furthermore, we discuss the practical implementation of our framework, including the design of attestation protocols, the selection of relevant model properties to verify, and the development of tools for generating and validating attestations. We illustrate the effectiveness of our approach through case studies in areas such as fairness in lending, transparency in healthcare, and safety in autonomous driving. Our results demonstrate the potential of auditable attestations to enhance trust and accountability in AI systems, fostering responsible innovation and deployment.
Third-party risk management (TPRM) reaches an inflection point, with artificial intelligence (AI) capabilities meeting pressing demands for real-time vendor risk oversight of increasingly complex digital ecosystems. Conventional assessment methodologies resting on manual questionnaires, annual review cycles, and document-centric evaluations are poorly matched to the pace and interconnectedness driving modern technology. This article analyzes how intelligent automation is remaking basic processes in vendor governance, from optimization of questionnaires through semantic modeling to predictive monitoring allowed through continuous data synthesis. Unstructured vendor control documentation is now parsed by natural language models to extract control metadata and produce risk assessments that must be validated, rather than created, by humans. Algorithmic integrity is tackled with multi-model verification architectures that employ parallel processing pipelines where ensemble methods quantify confidence levels and flag gaps in the vendor control environment for risk subject matter expert review. Brain-inspired computing principles underpin system design, with hierarchical feature extraction possible, along with adaptive learning from assessment outcomes. Technical debt becomes a critical governance factor, particularly in the context of data dependencies and configuration management across model lifecycles. Explainable artificial intelligence provides transparency that is vital to regulatory recognition, allowing risk officers to trace decision pathways and understand feature attributions underlying automated recommendations. Convergence of distributed ledger technology with intelligent risk systems unlocks opportunities for tamper-proof audit trails and privacy-preserving attestations in support of cross-organizational governance frameworks framed by emerging digital resilience mandates.
Cryptocurrency markets are difficult to model due to high volatility and multi-scale dynamics. This study investigates the directional predictability of crypto asset prices across multiple forecast horizons using Support Vector Machines (SVM). A daily Ethereum dataset (2018-2025), comprising candlesticks, technical indicators, and sentiment data, is used to predict upward or downward price movements from one to thirty days ahead. Model interpretability is achieved through SHAP, a popular XAI methodology, which quantifies feature contributions across various horizons. Results show that short-term forecasts approach random performance, while accuracy rises steadily with horizon length, peaking near 70% around the 24-day horizon. SHAP analysis reveals that short horizons rely on fast-reacting momentum indicators, whereas longer horizons emphasize slower, trend-following features. These findings highlight that medium-term price movements contain more structured information and demonstrate how explainable machine learning can uncover horizon-dependent dynamics in digital asset markets.
This paper introduces CryptoBench, the first expert-curated, dynamic benchmark designed to rigorously evaluate the real-world capabilities of Large Language Model (LLM) agents in the uniquely demanding and fast-paced cryptocurrency domain. Unlike general-purpose agent benchmarks for search and prediction, professional crypto analysis presents specific challenges: \emph{extreme time-sensitivity}, \emph{a highly adversarial information environment}, and the critical need to synthesize data from \emph{diverse, specialized sources}, such as on-chain intelligence platforms and real-time Decentralized Finance (DeFi) dashboards. CryptoBench thus serves as a much more challenging and valuable scenario for LLM agent assessment. To address these challenges, we constructed a live, dynamic benchmark featuring 50 questions per month, expertly designed by crypto-native professionals to mirror actual analyst workflows. These tasks are rigorously categorized within a four-quadrant system: Simple Retrieval, Complex Retrieval, Simple Prediction, and Complex Prediction. This granular categorization enables a precise assessment of an LLM agent's foundational data-gathering capabilities alongside its advanced analytical and forecasting skills. Our evaluation of ten LLMs, both directly and within an agentic framework, reveals a performance hierarchy and uncovers a failure mode. We observe a \textit{retrieval-prediction imbalance}, where many leading models, despite being proficient at data retrieval, demonstrate a pronounced weakness in tasks requiring predictive analysis. This highlights a problematic tendency for agents to appear factually grounded while lacking the deeper analytical capabilities to synthesize information.
This research introduces the Decentralized Finance (DeFi) TrustBoost Framework, which combines blockchain technology and Explainable AI to address challenges faced by lenders underwriting small business loan applications from low-wealth households. The framework is designed with a strong emphasis on fulfilling four crucial requirements of blockchain and AI systems: confidentiality, compliance with data protection laws, resistance to adversarial attacks, and compliance with regulatory audits. It presents a technique for tamper-proof auditing of automated AI decisions and a strategy for on-chain (inside-blockchain) and off-chain data storage to facilitate collaboration within and across financial organizations.
This paper unveils a pioneering modular framework for Decentralized Explainable Artificial Intelligence (DeXAI), harnessing blockchain to deliver unparalleled trust and clarity in AI systems. Addressing the opacity and privacy challenges of centralized AI, our framework integrates federated learning with Explainable AI (XAI) methods, namely SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), to produce intuitive explanations for AI decisions across distributed networks. A blockchain layer secures predictions and explanations, with smart contracts ensuring ethical compliance and auditable trails. Designed for adaptability, the framework supports diverse AI models and blockchain platforms, excelling in critical sectors like healthcare and finance. Our prototype validates its scalability and effectiveness, setting a new benchmark for trustworthy AI.
The Paris Journal on AI & Digital Ethics Bootstrapping Trust across Web2 and Web3 Domains Using Publicly Verifiable Web Data Yuan Lu¹, Qiang Tang² Corresponding authors:luyuan@iscas.ac.cn • qiang.tang@sydney.edu.au Abstract Through […]