Agentic, Context-Aware Risk Intelligence in the Internet of Value

Abstract: The Internet of Value (IoV) is a heterogeneous, partially-trusted network in which the dominant marginal risk is composite (route, sentiment, liquidity, and the policy a system is willing to commit to) rather than a property of any single chain. We argue that a risk primitive adequate for this regime is a composition of five engines: a prediction engine over price, liquidity, volatility, and route health; a Bittensor verification subnet that decentralises and economically scores prediction outputs; a sentiment-fusion engine over text, on-chain flow, and grey-literature feeds; an agentic engine under constitutional, role-bound action constraints; and an API-risk and scenario engine that converts forecasts into pre-committed action programs in the sense of Monte-Carlo scenario generation. We anchor the architecture in two empirical artefacts: a 27-hour policy-constrained liquidity stress-response experiment on Solana, and a 168-hour prediction-router calibration arc reported with explicit class-imbalance honesty. The case study supports deployability; the validator-loss decomposition is stated formally and is falsifiable.

• The paper presents a five-engine risk intelligence architecture unifying prediction, verification, sentiment fusion, agentic policy, and scenario planning to manage DeFi risks.
• It employs rigorous empirical methods, including live-API benchmarks and case studies, to validate safety circuits and ensure human-in-the-loop auditability.
• The design enhances decentralized governance through modular, killable components and incentive-compatible verification, setting new standards for multi-chain risk management.

Agentic, Context-Aware Risk Intelligence Architecture for the Internet of Value

Introduction and Motivation

The paper "Agentic, Context-Aware Risk Intelligence in the Internet of Value" (2605.05878) addresses systemic limitations of existing cross-chain risk assessment in decentralized finance (DeFi) ecosystems. Traditional risk engines are siloed within single chains and fail to account for compound risks manifest across bridge fragmentation, liquidity dispersion, narrative contagion, and agentic (LLM-mediated or automated) execution risk. The work introduces a new context-aware risk intelligence primitive for the Internet of Value (IoV), conceptualized as a composition of five engines over a shared observability and data ingestion substrate.

This five-engine architecture aims to unify multi-source data ingestion, decentralized prediction, sentiment fusion, agentic decision-making under constitutional constraints, and scenario-based pre-commitment into a deployable and auditable risk surface. The paper emphasizes honest benchmarking, empirical deployments, and the architectural principle that the verification substrate must itself be incentive-compatible and adversarially robust.

The Five-Engine OmniRisk Architecture

The proposed OmniRisk architecture comprises:

1. Prediction Engine: Generates probabilistic forecasts on price, liquidity, volatility, and route health at granular pool, route, and asset scales. The system processes over 66,000 samples across a 30-day horizon, with a baseline live-API accuracy around 53% on broad cohorts—highlighting the architectural emphasis on verification rather than marginal signal performance.
2. Bittensor Verification Subnet: Decentralizes prediction generation and verification over a Bittensor subnet, where miners propose predictive beliefs and validators score them against realized events using a convex combination of Brier error, forecast consistency, and explicit calibration. This mechanism draws from prediction-market game-theoretic principles and is architected for economic incentive alignment, robust adversarial tolerance, and regulatory kill-switch properties.
3. Sentiment-Fusion Engine: Aggregates and fuses multimodal, multi-source off-chain signals (news, social, grey literature) and on-chain telemetry using late-fusion and stacking architectures. This supports rapid detection of sentiment or narrative shocks that precede observable on-chain states.
4. Agentic Engine: Serves as a constitutional policy layer where actions are strictly constrained by declarative, role-specific constitutions. Every agentic action—performed via LLM or deterministic policy—is audited, cannot self-escalate privileges, and is gated behind explicit, auditable governance processes.
5. API-Risk and Scenario Engine: Converts predictive distributions into pre-committed, resource-bounded action programs. Actions are only triggered when realized scenarios cross pre-defined thresholds, consistent with Monte-Carlo scenario planning paradigms, all bounded by explicit governance and human-in-the-loop escalation layers.

The engines are composed over a canonical, time-aligned shared fabric, exposing unified access via gRPC, REST, GraphQL, and SDKs. The architecture is visualized and detailed in the following figures:

Figure 1: OmniRisk's five-engine architecture, with ingestion, agentic, prediction, verification, sentiment fusion, and scenario engines composed over a unified time-aligned intelligence fabric.

Figure 2: Dependency graph of the five-engine composition, where shared observability feeds parallel engines; only the agentic engine integrates and is governance-gated.

The architecture explicitly separates agent-task management from execution orchestration, mandating that runtime orchestration cannot bypass agentic policy boundaries—a constitutional AI constraint enforced in the system's runtime (LangGraph + Paperclip models).

Threat Model and Verification Specification

The paper formalizes a rigorous threat model, targeting robustness against:

• Malicious or colluding Bittensor miners/validators,
• Adversarial cross-venue manipulation, oracle attacks, and sentiment poisoning,
• Replay and governance bypass by compromised infrastructure actors.

For Bittensor subnet verification, the validator's loss for each miner is a tunable convex combination:

$$L_i = \alpha \cdot \text{Brier}(p_i, y) + \beta \cdot \text{Inconsistency}(m_i) + \gamma \cdot \text{Calibration}(c_i)$$

with $\alpha = 0.6$, $\beta = 0.2$, $\gamma = 0.2$. The system defines precise criteria for material chain-state events ($E_1$–$E_4$), including route-level liquidity contraction, price effect, bridge/oracle anomalies, and governance transitions. Importantly, the Bittensor subnet path is "killable"—if cross-miner variance in validation scores becomes excessive, the decentralized verification rail can be disabled in favor of fallbacks, bounding adverse impact.

Empirical Case Study: Policy-Constrained Liquidity Intervention

A focused empirical experiment demonstrates the architecture's deployability and the operability of constitutional policy boundaries. Over a 27-hour window, the scenario and agentic engines execute a "Trader" role policy in a Solana micro-cap pool (RYA/SOL):

• Phase I: During a 60%+ price collapse, policy-bound, time-sliced buys are executed with board approval and cap escalation.
• Phase II: Successive stress triggers governance interactions, with escalating caps and ratcheting ladder logic—each maneuver audited and externally referenced.
• Phase III: A catastrophic sell event triggers a 30% stop-loss, pausing the engine as designed; policy boundaries are preserved.

The case study demonstrates three key properties: precise safety circuit firing, full human-in-the-loop auditability, and strict role contract compliance, all verifiable from the public issue log and execution records.

Calibration Arc and Performance Metrics

A production-deployed calibration arc records 168-hour rolling metrics. Key numerical results:

• Live-API baseline accuracy: ~53% (broad cohort).
• Post-fix, concentrated cohort accuracy: up to 99.3% with Brier error at 0.1335 (top-1,000 assets by market cap).
• The result is strongly caveated: headlined accuracy is a product of low base-rate (class imbalance), with calibration error and cohort stratification being the true limiting factors. Brier error (0.1335) is closer to perfect than random but substantially above the perfect-scorer lower bound.

These results highlight both the importance of honest evaluation (class-imbalance caveats) and that the architecture's main innovation is its verification and composition, not pure predictive lift.

Architectural and Theoretical Implications

The proposed primitive offers several advances:

• Agentic Audibility: Deterministic and LLM-mediated agents are verifiably constrained, supportive of regulatory and audit needs for algorithmic governance in DeFi.
• Composability & Killability: Engines are killable, fusable, and updateable in isolation, supporting robust system modification and containment in adversarial scenarios.
• Verification as a Primary Citizen: Explicit incentive-compatible, modular, and composable verification surfaces are foregrounded over black-box accuracy claims.
• Empirical Honesty: Throughout, empirical results are honestly caveated with respect to class imbalance, implementational cohort restrictions, and deployment limitations.

This architecture generalizes to broader risk intelligence primitives in multi-agent, multi-chain, and dynamic adversarial environments relevant for scalable financial, identity, and governance protocols.

Limitations and Future Directions

The paper transparently enumerates limitations, all stated as explicit falsifiable claims:

• Outcomes are partially validated: sentiment-fusion and Bittensor subnet verification are not yet measured at production scale.
• Top-cap cohort restriction biases reported high accuracy; broad-cohort calibration claims remain unfalsified.
• The empirical case study is limited to a single Solana pool and a deterministic policy, not a full LLM-mediated agent.

Future work should conduct larger-scale, multi-pool, multi-chain deployments with adversarial test nets for the validator component, measure miner calibration properties under live Yuma market rules, and evaluate robustness to coordinated manipulation and sophisticated governance attacks.

Conclusion

The work defines a compositional, adversarially robust, and agentically auditable risk intelligence primitive for the Internet of Value. Grounded in both pragmatic deployments and rigorous empirical documentation, it sets a new standard for context-aware, verifiable risk management in heterogeneous and adversarial DeFi settings. Explicit architectural, adversarial, and empirical limitations are acknowledged, making the proposal both falsifiable and extensible. This architecture signals a transition towards constitutional, governance-bounded, and verifiably agentic AI systems in the decentralized economic domain.