Technical Summary: AURORA – An Uncertainty-Aware Resilience Micro-Agent for Causal Observability

1. Problem Statement

The computing continuum, which distributes computation across edge, fog, and cloud tiers, faces significant challenges in maintaining runtime resilience due to system heterogeneity. A primary obstacle is the prevalence of grey failures—subtle, non-catastrophic faults such as network congestion, CPU contention, and memory leaks. These failures produce ambiguous, overlapping symptoms (e.g., throughput degradation, latency spikes, SLO violations) that are difficult to distinguish from transient workload fluctuations.

Existing approaches often fail in this context for two reasons:

1. Lack of Causal Awareness: Rule-based systems and correlation-driven monitoring cannot reliably distinguish between competing fault hypotheses, leading to incorrect remediation (e.g., restarting a healthy service due to a network issue).
2. High Epistemic Uncertainty: Many active inference (AIF) frameworks act under high uncertainty without explicit safety bounds. In resource-constrained edge environments, acting on a poorly calibrated model can result in "destructive interventions" that exacerbate faults or propagate contention.

Current macro-level governance approaches struggle with real-time, autonomous triage on edge devices, while existing micro-agent frameworks often lack mathematical safety bounds to prevent action when epistemic uncertainty is dangerously high.

2. Methodology: The AURORA Framework

The paper proposes AURORA (Uncertainty-Aware Resilience Micro-Agent for Causal Observability), a lightweight framework designed for edge-tier environments. It integrates Active Inference (AIF), Bayesian Networks (BNs), and Pearl's do-calculus to enable localized causal reasoning and safe remediation.

Core Architectural Components

• Parallel Micro-Agent Pipeline:
  • Observation Layer: Four concurrent, single-responsibility micro-agents monitor specific system facets (Network Delay, Node Load, Camera Traffic, Bandwidth). They flag anomalies when specific SLO predicates ($\phi_i$) are violated.
  • Causal Inference: Upon detecting anomalies, a dedicated inference agent constructs a posterior distribution over candidate root causes ($C$) using a pre-specified Bayesian Network. To ensure tractability on resource-constrained hardware (e.g., Raspberry Pi), inference is restricted to the Markov Blanket of the symptom node, excluding conditionally independent variables.
  • Root Cause Ranking: Candidates are ranked using a composite score ($\rho$) that combines:
    1. Interventional Contrast ($\Delta$): The change in symptom probability under a counterfactual intervention ($do(C)$) versus a benign assignment.
    2. Observed Severity ($\sigma$): The magnitude of the anomaly.
    3. Structural Priors: Weights favoring direct parents of the symptom and root nodes over mediators.

The Dual-Gated Execution Mechanism

AURORA introduces a novel safety mechanism that authorizes local remediation only if two orthogonal conditions are met. If either fails, the agent abstains from local action and escalates the diagnostic payload to the fog tier.

1. Gate 1: Posterior Certainty Gate:

   • Checks if the maximum posterior probability ($P_{max}$) of the top-ranked root cause exceeds a dynamic threshold ($\tau$).
   • Dynamic Scaling: The threshold $\tau$ is relaxed linearly based on the number of concurrent anomalies ($n$) to prevent total abstention during multi-fault scenarios, bounded by a minimum floor ($\tau_{min}$).
   • Purpose: Ensures the agent is confident enough to distinguish the fault class.

2. Gate 2: Variational Free Energy (VFE) Safety Gate:

   • Checks if the predicted post-intervention state is consistent with current observations. It decomposes VFE ($F$) into:
     • Pragmatic Term ($F_{pragmatic}$): Penalizes actions with low predicted SLO recovery probability.
     • Epistemic Term ($F_{epistemic}$): Penalizes actions that destabilize the model's posterior distribution across the network.
   • Purpose: Ensures the generative model is calibrated and the proposed action will not produce "surprise" (inconsistency) in the system state.

Remediation Logic

If both gates pass, the agent executes the action ($a^*$) that maximizes the probability of SLO recovery ($P(S=1 | do(a))$). If either gate fails, the system engages intentional abstention, serializing the diagnostic payload and offloading it to the fog tier for deeper analysis, rather than risking a destructive local intervention.

3. Key Contributions

The paper summarizes three primary technical contributions:

1. Lightweight Parallel Architecture: A micro-agent design that performs localized causal observability and AIF-based recovery reasoning under posterior fault belief and epistemic uncertainty, suitable for edge devices.
2. Bounded-Complexity Causal Modeling: The use of Bayesian Networks with Markov blanket-constrained inference to enable real-time counterfactual root-cause diagnosis via do-calculus without exhaustive system-wide state evaluation.
3. Dual-Gated Safety Mechanism: A novel execution protocol that authorizes local recovery only when causal confidence is high and predicted epistemic uncertainty is bounded, otherwise escalating to the fog tier.

4. Experimental Results

The framework was evaluated in a Python-based simulator emulating a resource-constrained video-streaming edge device (Raspberry Pi 4) under 30,006 Monte Carlo trials involving three grey-fault classes: Network Bandwidth issues, Anomalous Processing (CPU spikes), and Memory Exhaustion.

Comparative Baselines:

• Rule-Based Agent: Static thresholds with no causal reasoning.
• AIF (No Gate) Agent: Standard AIF with BNs but no epistemic safety bounds.

Key Metrics:

• Destructive Action Rate: AURORA achieved 0.0% destructive actions across all trials. In contrast, the Rule-Based and AIF (No Gate) baselines exhibited a 33.3% destructive action rate.
• Repair Accuracy: AURORA maintained a 62.0% repair accuracy. While lower than the 66.7% of the ungated AIF baseline, this is attributed to the framework's refusal to act under high uncertainty (safety-first design).
• Abstention Rate: AURORA abstained in 65.9% of cases, acting as a safety buffer. The dual-gate mechanism was non-redundant: the Posterior Certainty Gate triggered abstention in 44.3% of cases, while the VFE Gate triggered it in 21.6%, with overlap accounting for the total abstention rate.
• Mean Time to Repair (MTTR): Despite the computational overhead of Bayesian inference and VFE calculation, AURORA achieved an average MTTR of 3ms, confirming feasibility for real-time edge deployment.

5. Significance and Claims

The paper claims that AURORA demonstrates the feasibility of uncertainty-aware micro-agent resilience at the edge. Its primary significance lies in shifting the paradigm from "forced autonomy" to "intentional abstention."

The authors argue that in unpredictable grey-failure conditions, recognizing and bounding uncertainty is more valuable than attempting to resolve every fault autonomously. By strictly enforcing safety bounds, AURORA eliminates the risk of destructive cascading failures (common in rule-based and ungated AIF systems) while maintaining diagnostic accuracy comparable to unbounded models. The framework achieves a "graceful degradation of autonomy," where the edge tier handles confident, low-uncertainty cases locally, and escalates ambiguous, high-uncertainty cases to the fog tier, rather than making blind, potentially harmful decisions.

The paper concludes that while AURORA successfully prevents destructive interventions, future work is required to deploy the system on physical edge-fog testbeds to evaluate performance under realistic physical constraints (e.g., telemetry noise, actuation delays) and to automate the calibration of SLO predicates and gate thresholds.