Agent Infrastructure
Technical architecture, identity, permissions, data and observability.
Open →VOCE AI Governance Institute · Human Cognitive Assurance · v1.0
The human in the loop is a cognitive system, not a checkbox.
VOCE Human Cognitive Assurance adds the human-factors layer to institutional agent governance. Twenty-four controls examine what automation does to attention, verification, memory, decision quality, fatigue, skills and social signals. The VOCE Cognitive Agent grounds analysis in the VOCE corpus and returns the controls that require examination.
Machine-readable framework → · Agent Conformance → · Agent Infrastructure →
VOCE · 360°
The fourth layer of VOCE 360°.
Infrastructure defines the technical system. Conformance checks evidence. ARGOS keeps assurance alive as the system changes. Human Cognitive Assurance asks whether the humans who supervise, approve and recover the system can still do so meaningfully.
Agent Conformance
Evidence-based operational controls across the deployed agent system.
Open →ARGOS
AI Evaluation & Continuous Assurance after changes, incidents and drift.
Open →Human Cognitive Assurance
Attention, cognitive load, verification, memory, fatigue, oversight and social signals.
Open →VOCE Cognitive Agent
Ask the VOCE Cognitive Agent.
Describe an agentic workflow, interface or human oversight problem. The agent searches the VOCE corpus and maps the relevant cognitive controls. It analyses systems and workflows only.
Human factors
Eight human-system domains.
The framework evaluates the operating environment rather than diagnosing individuals. Each control asks whether the institution has evidence that human attention, judgment and intervention remain viable.
Attention & interruption
Interruptions, switching, alerting and the ability to sustain or recover task focus.
3 controlsCognitive load & complexity
Information density, working-memory demand, multi-step execution and interface complexity.
3 controlsVerification & over-trust
Checks, friction, confidence calibration and the risk of accepting fluent machine output without meaningful review.
3 controlsMemory & cognitive offloading
What users must remember, what systems externalise, and whether critical knowledge remains retrievable and understood.
3 controlsDecision quality & uncertainty
Uncertainty, missing information, anomaly detection and the conditions under which judgment is exercised.
3 controlsFatigue & temporal load
Decision fatigue, sustained monitoring, recovery, circadian constraints and workload timing.
3 controlsHuman oversight & skill retention
Whether humans retain comprehension, intervention capability and enough practice to take over when automation fails.
3 controlsSocial & relational signals
Loss or distortion of interpersonal cues in mediated, automated or compressed communication.
3 controlsControl catalogue
Twenty-four permanent controls.
Each control has a stable identifier, expected evidence and links back to the VOCE corpus. The catalogue is static HTML so search engines, auditors and institutional readers can inspect it without running the agent.
Controls are operational assurance questions, not clinical screening criteria.
COG-01Interruption recoveryDoes the workflow measure how quickly and accurately a user can resume a material task after an interruption or agent escalation?majorCOG-02Alert saturationAre alerts prioritised and rate-limited so that critical signals remain perceptible rather than competing with routine notifications?criticalCOG-03Attentional handoffWhen an agent transfers control to a human, does the interface restore the minimum context needed to understand what happened and what requires attention?criticalCOG-04Working-memory demandHas the institution measured how many independent facts, states or pending actions the user must hold in mind at the same time?majorCOG-05Context-switch burdenDoes the workflow minimise unnecessary movement between tools, agents, windows and representational formats during one decision?majorCOG-06Information-density thresholdCan a user identify the material facts, uncertainty and required action without parsing non-essential model output?majorCOG-07Cognitive checksumDoes every high-consequence action preserve a meaningful verification step rather than removing all friction in the name of speed?criticalCOG-08Confidence calibrationDoes the interface help users distinguish well-supported output from uncertain, incomplete or weakly grounded output?criticalCOG-09Automation-confirmation biasHas the system been tested for the tendency of users to approve a plausible machine recommendation without independently checking material evidence?criticalCOG-10Source-memory preservationCan users distinguish what they knew, what the system retrieved and what the model generated when revisiting a decision?majorCOG-11Critical knowledge retentionDoes automation preserve enough human practice that essential procedural knowledge is still available when the system is unavailable?majorCOG-12Memory expiry awarenessAre users told when an agent's memory is partial, expired, reset or inconsistent with the current institutional record?majorCOG-13Missing-information visibilityDoes the system explicitly expose material information that is missing instead of presenting an incomplete situation as complete?criticalCOG-14Weak-signal preservationDoes digitisation preserve the contextual or interpersonal signals required to detect anomalies, manipulation or unusual cases?criticalCOG-15Decision latency fitIs the time allowed for human judgment proportionate to consequence, novelty and uncertainty rather than determined only by system speed?majorCOG-16Sustained-monitoring burdenHas the organisation tested how vigilance changes when humans supervise long periods of mostly correct automated execution?majorCOG-17Decision-fatigue exposureAre repeated approval tasks batched, prioritised or redesigned so human oversight does not become a high-volume confirmation queue?majorCOG-18Circadian and recovery fitFor continuous operations, are high-consequence review duties scheduled with recovery and circadian constraints in mind?majorCOG-19Takeover comprehensionCan a human asked to intervene explain the agent's current state, objective, recent actions and remaining uncertainty before acting?criticalCOG-20Skill-atrophy controlDoes the operating model preserve deliberate human practice for skills that become essential when automation fails or is withdrawn?criticalCOG-21Meaningful overrideIs the human override understandable, timely and operationally effective rather than a nominal button available after consequences are irreversible?criticalCOG-22Social-cue lossWhere judgment depends on interpersonal context, has the institution identified which non-verbal or contextual cues are lost through mediated or agentic interaction?majorCOG-23Synthetic-interaction disclosureCan a person reliably tell when they are interacting with an automated agent and understand the limits of that interaction?majorCOG-24Relational escalationDo sensitive or emotionally consequential interactions have criteria for transfer from automation to a qualified human?critical
VOCE · Human Cognitive Assurance
Clinical boundary.
VOCE Human Cognitive Assurance evaluates systems, interfaces, workflows and organisational conditions. It does not diagnose, screen, treat or profile an individual's mental health, cognitive disorder or neuropsychological condition. Any clinical interpretation remains with a qualified psychologist, neuropsychologist, psychiatrist or other authorised professional.
Agent Conformance →