EVIDENCE TRAIL
Adaptive workload balancing — route reviews by reviewer fatigue
Verbatim excerpts from the upstream sources cited on the mitigation page, with what each source does and does not prove. The strongest upstream match is OWASP Agentic AI v1.1 §T10, which names "adaptive workload distribution across human reviewers" verbatim. Note: the MDX cites NIST AI 600-1 MANAGE-4.2 for "human-reviewer workload management" — the actual MANAGE-4.2 actions cover continual improvement processes, not workload routing specifically; see that entry's notSupports for the corrected framing.
Last cross-checked against upstream sources: · 7 sources
References
Each entry shows what the source supports and what it does not prove.
OWASP Agentic AI — Threats & Mitigations v1.1
§T10 Overwhelming Human-in-the-Loop — Threat Description
"Overwhelming Human-in-the-Loop (HITL) occurs when attackers exploit human oversight dependencies in multi-agent AI systems, overwhelming users with excessive intervention requests, decision fatigue, or cognitive overload. This vulnerability arises in scalable AI architectures, where human capacity cannot keep up with multi-agent operations, leading to rushed approvals, reduced scrutiny, and systemic decision failures."
Supports: Names decision fatigue and cognitive overload as the primary attack mechanism against HITL systems. Establishes that workload saturation is an exploitable vulnerability, not merely an operational nuisance — the threat framing that makes this mitigation necessary.
Does not prove: Does not prescribe workload-routing or fatigue-signal thresholds as the defence. The description names the problem; the mitigation section (see next entries) supplies the defence.
OWASP Agentic AI — Threats & Mitigations v1.1
§T10 Overwhelming Human-in-the-Loop — Mitigation Steps (Step 1: Proactive controls)
"Apply adaptive workload distribution across human reviewers. Balance AI review tasks dynamically to prevent decision fatigue for individual reviewers."
Supports: Verbatim upstream source for the adaptive-routing pattern this control implements. The phrase "adaptive workload distribution" is the closest upstream match for the control's title. Names the explicit goal (prevent decision fatigue for individual reviewers) and the mechanism (dynamic balancing of review tasks across the pool).
Does not prove: Does not specify what fatigue signals to measure (reviews-per-hour, time-on-task, agreement-decay), nor the routing algorithm. Helmwart adds the signal-composition and mandatory-break mechanics that upstream leaves undefined.
OWASP Agentic AI — Threats & Mitigations v1.1
§T10 Overwhelming Human-in-the-Loop — Mitigation Steps (Step 1: Proactive controls)
"Use AI trust scoring to prioritize HITL review queues based on risk level. Automate low-risk approvals while requiring human oversight for high-impact tasks. Limit AI-generated notifications to prevent cognitive overload. Implement frequency thresholds to limit excessive AI-generated notifications, requests, and approvals to prevent decision fatigue."
Supports: Establishes risk-stratified queue prioritisation and frequency thresholds as upstream-recommended controls for decision fatigue. Directly supports pairing this control with m-risk-prioritized-queue and m-fail-closed.
Does not prove: Trust-scoring here refers to AI-generated risk levels, not reviewer fatigue scoring. The passage does not address the reviewer-side signal composites that drive Helmwart's routing algorithm.
OWASP Agentic AI — Threats & Mitigations v1.1
§T10 Overwhelming Human-in-the-Loop — Threat Summary Table (Mitigations column)
"Develop advanced human-AI interaction frameworks, and adaptive trust mechanisms. These are dynamic AI governance models that employ dynamic intervention thresholds to adjust the level of human oversight and automation based on risk, confidence, and context. Apply hierarchical AI-human collaboration where low-risk decisions are automated, and human intervention is prioritized for high-risk anomalies."
Supports: Names dynamic intervention thresholds and hierarchical human-AI collaboration as the governance model for T10. Supports the escalation and queue-deprioritisation paths in this control.
Does not prove: Framing is about risk/confidence of the AI decision, not workload/fatigue of the reviewer. Complementary rationale rather than a direct statement of workload-routing.
MITRE ATLAS AML.M0029 — Human In-the-Loop for AI Agent Actions
AML.M0029 mitigation description — paragraph 2
"Human In-the-Loop policies should follow the degree of consequence of the task at hand. Minor, repetitive tasks performed by agents accessing basic tools may only require minimal human oversight, while agents employed in systems with significant consequences may necessitate approval from multiple stakeholders diversified across multiple organizations."
Supports: Establishes consequence-proportionate HITL as a MITRE-endorsed pattern. The "multiple stakeholders diversified across multiple organizations" framing supports the reviewer-pool distribution mechanic. The consequence-scaling logic is the precondition that makes adaptive routing meaningful.
Does not prove: Does not address workload, fatigue, or routing among reviewers within a single queue. AML.M0029 defines when human review is needed; this control defines how to distribute that review work without degrading decision quality.
NIST AI 600-1 — Generative AI Profile (NIST AI RMF)
MANAGE 4.2 — "Measurable activities for continual improvements are integrated into AI system updates and include regular engagement with interested parties"
No verbatim excerpt pulled yet — open the original to verify the cited section.
Supports: MANAGE 4.2 mandates continual improvement measurement for deployed AI systems. When reviewer-fatigue metrics (agreement-rate decay, decision-reversal rate) are operationalised as quality signals, this subcategory becomes the governance hook for tracking and acting on them.
Does not prove: MANAGE 4.2's actions (MG-4.2-001 through MG-4.2-003) cover performance reporting, incident post-mortems, and output visualisation — none explicitly names human-reviewer workload management or fatigue-routing. The MDX's claim that MANAGE-4.2 "names human-reviewer workload management" is an overstatement; the correct framing is that MANAGE-4.2 provides the continual-improvement process under which fatigue metrics would be tracked. The closest specific NIST action is MG-3.2-008, which names "human moderation systems" and human-AI configuration policies.
OWASP Agentic AI Mitigation Playbook P5 (HITL Decision Fatigue)
No verbatim excerpt pulled yet — open the original to verify the cited section.
Supports: Internal Helmwart mapping of the OWASP P5 playbook, which covers the protective-phase controls for HITL decision fatigue including adaptive review routing. Names this mitigation directly in the protective-phase sequence.
Does not prove: Internal page, not the source of the workload-routing principle itself.