In the framework of the NeoMundi AI Observatory’s work, we are publishing today a contribution by James Moore, founder of Nova Jema AI Systems and specialist in the governance of artificial intelligence systems under real-world execution conditions. He kindly shared with us a precise reflection on a frequently overlooked subject: once we measure an AI’s behavior, how can we ensure that a critical signal does not turn into an action without a legitimate human authority having been clearly identified and activated?
Human & Execution-Time Governance Contributor
An AI can produce a response that is stable or unstable, consistent or variable, factually solid or fragile, or difficult to interpret. Organizations now have tools to detect these signals: drift, unusual variation, factual risk, coherence breaks, or insufficient measurement coverage. This observation capability represents real progress. It makes visible what would otherwise remain invisible.
However, an essential question often remains without a clear answer: when a signal appears, who has the right to act? Who bears the responsibility if the action continues? Who can suspend, escalate, or stop the process?
It is precisely at this frontier that operational AI governance begins.
A Signal Is Not an Authority
NeoMundi continuously observes the behavior of AI systems in production: their stability, variability, risk signals, coherence, validity, cost, and execution conditions. This measurement is essential because it makes readable what would otherwise remain silent. However, a risk signal does not decide anything by itself.
An alert does not know who is authorized to intervene, nor who carries final responsibility for a decision. It also does not know whether a response can be sent to a client, whether an action must be suspended, whether human escalation is mandatory, or whether the context has evolved since the last validation.
This is the heart of James Moore’s contribution:
Runtime signals do not constitute an authority. They provide evidence that must then enter a governance chain.
In other words, an AI can be observed with precision. But observation alone must never be sufficient to authorize an action.
The Most Frequent Risk: Seeing the Problem Without Being Able to Intervene
Consider the case of an AI that responds to clients, handles cases, recommends actions, or synthesizes sensitive information. A measurement instrument detects an increase in factual risk or unusual instability. The organization then finds itself in a paradoxical situation: it has an alert, it sees that a problem potentially exists, but no role is clearly designated to decide, no escalation procedure is activated, and no one really knows whether to continue, suspend, or stop.
This gap creates a form of visibility without intervention capability. In environments where decisions have real consequences, this disconnect between detection and authority becomes a major operational risk.
From Measurement to Action: A Simple Chain
James Moore’s proposal consists of clearly distinguishing the steps between technical detection and the authority decision.
Runtime signal observed
↓
Signal interpretation
↓
Authority verification
↓
Accountable role identified
↓
Escalation, human review, continuation, or halt
↓
Verifiable decision record
This separation is essential. Measurement answers a question of evidence: what was observed? Governance answers a question of authority: who can decide, under what responsibility, and within what limits?
These two functions must collaborate, but they must not be confused.
The Runtime Governance Contract: Making Responsibility Explicit
To make this critical transition operational, NeoMundi has launched an open contribution process. James Moore’s structured proposal, Runtime Governance Contract v0.1 (PDF available below), provides a strong initial foundation for this work. This contract does not replace internal policies, regulatory obligations, or human judgment. Its objective is more targeted: to prevent a measured signal from silently turning into an authorized action.
The governance record associates each signal with several essential elements: its identifier, the time of detection, the relevant context, the risk and confidence levels, the intended action, the required authority, the responsible role, the escalation path, the permitted and prohibited actions, as well as the final status and the audit trail required.
The objective is to make visible what should already exist in any organization that entrusts important decisions to AI: who decides, who is accountable, what is permitted, and what must stop?
A Concrete Example: An AI Answer Intended for a Client
Imagine an AI preparing an explanation for a client in a regulated sector (banking, insurance, healthcare, public service, or human resources management). The measurement layer detects a factual risk signal.
The appropriate reflex is no longer to send the response automatically because “the AI generated it.” Instead, the signal is recorded, the intended action is identified, a human review is required, the compliance officer is designated, and direct sending to the client is blocked. The final decision is then recorded with its justification.
The measurement does not declare that the text is legally incorrect. It simply indicates that the observed conditions require verification before this response becomes a real action.
This distinction profoundly changes the way the organization maintains control.
Human Presence Is Not Always Enough
Many systems today display the promise “a human remains in the loop.” This formulation alone guarantees nothing. A human can see an alert without having the real power to interrupt the flow, without the mandate to validate or reject, without the necessary time, without sufficient information, without clearly assigned responsibility, or without a precise rule that triggers escalation.
A human who is present but lacks real authority becomes decorative supervision. Execution-time governance requires more: a named role, defined rights, identifiable responsibility, and an effective ability to suspend or stop an action.
What This Contribution Adds to NeoMundi
NeoMundi develops tools for the continuous measurement of AI behavior. James Moore’s contribution adds an indispensable complementary step: how to link a runtime signal to a legitimate, traceable, and proportionate human decision.
This articulation is particularly critical in environments where an AI response can have a direct impact on a client, citizen, patient, employee, financial file, insurance decision, industrial operation, or critical infrastructure.
NeoMundi does not claim to replace human governance. Its objective is to provide an independent measurement layer capable of clearly indicating what was observed, under what conditions, and at what moment, so that the organization can then connect this observation to its own rules of authority, escalation, and responsibility.
An Open Basis for Discussion
The Runtime Governance Contract v0.1 constitutes a first proposal. It is neither a finalized standard, nor a certification, nor a sufficient legal instrument in itself. It is an open working basis intended for researchers, DPOs, risk managers, architects, lawyers, infrastructure operators, and AI teams.
The question remains deliberately simple: how to prevent a technical signal from turning, without explicit human control, into a decision that engages an organization or affects a person?
It is both an engineering question and a question of responsibility.
Conclusion
An AI can generate a response. An instrument can measure its behavior. But no measurement should, by itself, become an authority. Measurement makes a situation visible. Governance determines whether the action remains admissible. It is at this frontier between signal, authority, responsibility, and execution, that a crucial part of trust in tomorrow’s artificial intelligence systems will be decided.
About the Contribution
This contribution was prepared by James Moore, Founder of Nova Jema AI Systems, an independent AI governance research initiative focused on execution-time accountability, human authority structures, escalation boundaries, and operational governance gaps.
Download the contribution components, PDF:
Author
James Moore
NeoMundi Research Contributor · Execution-Time Human Governance
Execution-Time Governance, Operational Responsibility & Escalation
James Moore works on AI governance at the moment of execution: the layer that determines whether AI-assisted actions should be authorized, escalated, interrupted, or stopped before an irreversible consequence occurs.
His work focuses on the critical moment when an AI-produced or AI-assisted output becomes a real-world action. This is where authority, responsibility, escalation pathways, and operational admissibility must be structurally defined, particularly in high-risk or regulated environments.
Within the Observatory, he contributes to the human and operational layer of governance: the articulation between runtime signals and decision rights, escalation logics, responsibility structures, and secure intervention mechanisms.
Mission: To contribute to the definition of escalation boundaries, responsibility frameworks, human authority, and operational admissibility in high-stakes AI use cases.
Profile: [LinkedIn]
Thank you, Sébastien, for taking the time to develop and publish this contribution.
One of the key ideas I hoped to explore is that measuring AI behaviour and governing AI-assisted action are complementary, but fundamentally different, problems. Runtime signals provide evidence; they do not, by themselves, confer authority to act.
I’m grateful to NeoMundi for providing a space where these governance questions can be discussed openly, challenged constructively, and refined alongside measurement and observability research.
I look forward to continuing the conversation and learning from the community.