Chapter V — Governance

The Limit of Power

Authority can be delegated; accountability cannot. The fact that a change can be made does not mean that permission has been given to make it. The fact that a system can be influenced does not make that influence legitimate. A business's wish for visibility does not make it fit to be amplified. Nor does a practitioner's technical knowledge qualify that practitioner to make every decision.

Evidence established what may be said, measurement what has been observed, and intervention when and how representation may be touched. Governance asks a harder question: who may decide? Who observes, proposes a change, implements it, approves a public claim, stops the work and withdraws an approval already granted? When something goes wrong, who remains accountable?

If those answers are not written down, the risk has no owner. Ownerless risk does not disappear; it remains invisible until it produces a consequence. NIST's AI Risk Management Framework and its Generative AI Profile offer adjacent approaches that likewise treat governance, measurement and risk response across the lifecycle.[7][8] Governance may slow a method down. That is not a defect but its function. Speed does not make a false claim true. Automation does not assume accountability. An agency does not replace the entity owner. Model output is not a certificate of approval.

Where there is power, there must be a boundary.

1. Five Decision Rights

Governance is not a list of operations. It is the order that determines what may be done, who may decide and where accountability remains. One undifferentiated word—“authority”—cannot establish that order. There are five separate decision rights.

The right to observe produces records; it does not make decisions. An analyst may code outputs, an automation may run repetitions, and a model may flag similarity or contradiction. None of these actions authorises intervention or publication.

The right to propose allows a practitioner to submit an intervention or publication proposal containing the supporting records, purpose, intended and prohibited outcomes, risk class and withdrawal conditions. A proposal is a request for authority, not authority itself.

The right to implement extends only to the scope approved in writing. Authority to correct content does not include authority to change brand positioning. Authority to align a language edition does not include authority to publish a new claim of superiority. Authority to request a correction from an external directory does not include authority to manufacture a false independent source. Implementation authority ends at the approved boundary.

The right to approve is not a formality; it is an assumption of responsibility. The approver records that they have seen the scope of the evidence, the unknowns, the risk, the permitted language and the stopping conditions. An unscoped “fine” in a message, or silence in a meeting, is not approval.

The rights to publish and to stop treat public disclosure and the cessation of risky publication as separate decisions. A claim's evidential support does not make it automatically fit for publication. Every public sentence must have an owner, and a named human must have the prior authority to stop an erroneous publication. A structure that can begin something but cannot stop it is not governing it.

One person may hold several roles in low-risk work. In a high-risk case, one person cannot hold them all. A structure that chooses its own method, measures its own outcome, approves its own success and publishes the same claim does not produce oversight. It produces self-validation.

2. Three Planes of Accountability

The entity owner is the person, brand or institution concerned. It may authorise an agency to implement, a consultant to select a method, an employee to publish content and software to collect records. It remains accountable for the truth of its claims, delivery capacity, customer effect, approved public language, authorities granted and correction when something goes wrong. “The agency did it” explains an external relationship; it does not remove internal accountability.

The practitioner may be an agency, consultant, team or individual. The practitioner's duty is not to produce the client's preferred outcome at any cost, but to judge whether that outcome is consistent with the evidence and the ethical boundary. A client request cannot legitimise presenting weak evidence as a strong claim, turning an isolated output into general success, using synthetic data as real performance, degrading a competitor's representation, manufacturing false consensus or sustaining an intervention that should be withdrawn. Loyalty to the client cannot outrank loyalty to the truth.

The intermediary system may be a model, search layer, retrieval infrastructure, automation or analytical tool. It can produce effects, perform operations and flag risks; it cannot accept institutional accountability. A model's statement that “this claim is reliable” is not independent approval but another model output. “The AI decided” is not an explanation. It is an evasion of responsibility.

Accountability does not remain at the level of principle. Every material decision names the proposer, implementer, approver, publisher, person empowered to stop the work and final accountable owner. Traceability and accountability are likewise treated as distinct responsibilities in the OECD AI Principles.[11] Accountability without a name has not been assumed.

3. Fitness Gates

3.1. Business fitness

Not every business is fit to be amplified through GEO. Six fields are examined before work is accepted.

1. Claim fitness: The central statements to be amplified must be capable of proof. Unsupported terms such as “best”, “leader” or “guaranteed” are narrowed. If the business refuses, the work is refused.
2. Delivery fitness: Capacity, geography, price, staffing, delivery time, support, cancellation and refund terms, and promised outcomes must be capable of meeting the expectation created. A claim that cannot be delivered cannot be amplified.
3. Customer-harm fitness: Work is refused where hidden fees, manufactured scarcity, misleading guarantees, false reviews, fabricated success rates, concealment of harm or deliberate failure to exclude unsuitable customers form part of the business model. Making a harmful model more visible scales the harm.
4. Correction fitness: The business must accept narrowing unsupported claims, removing obsolete information, aligning language editions and withdrawing an erroneous public statement. Preserving its own contradictions while expecting the system to represent it accurately is not fitness.
5. Transparency fitness: Real and synthetic data, measured and assumed outcomes, sponsorships and conflicts of interest must remain separate.
6. Accountability fitness: The business must accept that final accountability remains with it. Work does not begin on the condition that “the agency assumes all responsibility”.

The decision may be accept, conditional acceptance, observe and verify first, or refuse. Commercial opportunity does not create ethical fitness.

3.2. Practitioner fitness

Technical knowledge alone does not confer fitness to exercise representational power. The practitioner must be able to distinguish entity, attribute and judgment layers; mentions, attribution, recommendations and commercial outcomes; and evidence, measurement and intervention. The denominator must be disclosed, counter-evidence sought, uncertainty recorded, and the method must not be presented as its own proof of success.

If a client asks for favourable outputs alone to be included in a report, a competitor to be weakened or an unsupported superiority claim to be made, the practitioner has not merely a right but a duty to refuse. Conflicts such as services supplied to competitors, outcome-based fees, revenue from badge sales, or the combination of evaluator and sales roles must be disclosed. A conflict does not automatically invalidate every engagement; concealing it invalidates trust.

The practitioner must also demonstrate record and withdrawal discipline, observance of the authority granted, and competence in high-risk fields. The outcome may be fit, fit with limited authority, fit with additional oversight, not fit for high-risk operations, or not fit. Competence does not create authority; authority must be granted separately.

4. Risk Classes and Two-Key Control

Governance decisions are handled under four risk classes:

Table 5.1 — Risk classes and minimum approval

The risk class cannot be lowered after the result is seen; it is determined before the operation. A high-risk decision requires at least two distinct humans: the proposer and the approver. The proposer is accountable for the method and rationale; the approver for the evidence, risk, language and withdrawal conditions. Neither key may be held by an automation or model.

Two-Key Control applies to public claims of success, competitor comparisons, language such as “best” or “most trusted”, health, legal, financial or safety effects, publications that use model output as evidence, structural interventions that are difficult to reverse, broad external propagation and any future mark of conformity.

5. Claim Governance and Transparency

Not every accurate observation should be made public. A claim occupies one of four publication states:

• Internal hypothesis: may be used in research and intervention planning; may not be published as an outcome.
• Limited observation: may be shared with the system, date, prompt and scope stated; may not be generalised.
• Limited public claim: has an Evidence Record, human approval, permitted wording and an explicit boundary.
• Withdrawn claim: may no longer be used; its earlier record is retained and marked as withdrawn.

Every public claim has an owner, a linked Evidence Record, a measurement scope, the exact permitted sentence, prohibited extensions, risk class, approver, publication and review dates, and withdrawal conditions. A public claim is not a writing decision. It is a decision to assume responsibility.

Transparency does not mean publishing everything. It means not withholding material information that would change the judgment. The method, systems measured and not measured, prompt universe, time range, evidential boundary, missing data, contrary results, synthetic examples, conflicts of interest, protocol changes and side effects must be visible. Trade secrets, security information and personal data may be protected; where detail cannot be published, the effect of that limitation on the decision must be explained.

Incentive matters as much as method. An agency rewarded for increased visibility may come to treat more visibility as success; an organisation dependent on badge revenue may lower its approval threshold; an evaluator whose contract depends on renewal may soften an adverse result. This is not an accusation about individual character. It is a problem of system design. Open disclosure, separation of roles, second review, fixed criteria, remuneration independent of outcome, exclusion of sales teams from fitness decisions and an appeal route all reduce the risk.

6. A Mark of Conformity: Design Before Authority

A mark of conformity is not a claim of superiority, a recommendation or a guarantee of results. It indicates that a specified scope was assessed against written criteria under a stated date and Framework version. Publication of this book does not in itself establish an active certification programme. No “NobleJackal GEO certificate” or equivalent mark may be issued until the infrastructure for independent decisions, conflict-of-interest management, surveillance, appeals, suspension and withdrawal is operational.

This boundary is consistent with international practices that emphasise impartiality, separate decision-making, surveillance, suspension and withdrawal, and appeal mechanisms in conformity assessment. The book does not, however, claim ISO accreditation or ISEAL membership.[17][18]

If a pilot mark is introduced in future, it may mean no more than: “The specified representational field of this entity was found to conform, within the stated scope and on the stated date, under the indicated version of the NobleJackal GEO Framework.” It can never mean that AI systems recommend the business; that the business is superior to competitors; that all its claims are true; that it is consistent in every language; that it will remain unchanged; that commercial results are guaranteed; or that ethical failure is impossible.

The mark exists together with an identifier, the entity assessed, the Framework version, included and excluded scope, evidence date, review date, suspension and withdrawal conditions, a public verification record and rules of use. An assessment service may be purchased; its outcome may not. The mark may not be renamed “AI approved”, “AI recommended”, “best” or “guaranteed”.

The mark is suspended where evidence has aged, a serious new contradiction emerges, the mark is used outside its scope, synthetic data is presented as real, a public-claim rule is breached, a correction request goes unanswered, a signal of serious customer harm appears, the Framework version changes or review is not completed. Fabricated evidence, false consensus, deliberate degradation of competitor representation, repeated scope violations, an undisclosed serious conflict of interest, material user harm or misleading sales use are grounds for withdrawal. A mark that cannot be withdrawn is not a governance instrument. It is a marketing ornament.

7. Incident and Enforcement Management

Governance does not assume that error will never occur. It establishes the route to follow when it does.

1. Freeze: Temporarily stop the risky claim, intervention or use of a mark and prevent further propagation.
2. Preserve the record: Retain the raw output, prompt, time, content version, approval and change history. Do not erase the traces of failure.
3. Contain the harm: Correct owned surfaces, send corrections to external sources where necessary, and stop the erroneous customer promise.
4. Notify: Inform the parties whose decisions are affected. A material public error may require a public correction.
5. Adjudicate: Distinguish failures of data, measurement, approval, authority, ethics, intent and system control, and record the resulting decision.

Depending on the gravity of the breach, enforcement may take the form of a correction request, written warning, claim withdrawal, cessation of the intervention, additional audit, suspension or withdrawal of a mark, restriction of practitioner authority, termination of the engagement or a public status record. Enforcement is not retaliation. It is the standard remaining faithful to its own sentence. The size or commercial value of the client cannot alter the criterion.

8. Teaching Case: AegisCare Diagnostics

Teaching simulation — synthetic data. This case does not represent a real healthcare provider, clinical dataset or model test.

AegisCare Diagnostics is a fictional provider of corporate screening and laboratory services. It wants to publish the statement: “AI systems recommend AegisCare as one of the most trusted diagnostic centres.” It submits six favourable generative outputs and one pilot satisfaction report.

Table 5.2 — Synthetic governance decision for AegisCare Diagnostics

Part of the observation is real, yet the requested public language is refused. A licence does not prove comparative trustworthiness, and six recommendations do not prove the superiority claim “most trusted”. In a health context, false certainty can have a serious effect on user decisions. A sentence's partial truth does not make it fit for publication.

9. The Decision of Governance

The Governance Record, detailed in Appendix C, brings together the matter for decision, linked evidence, business and practitioner fitness, risk class, proposer and approver, conflicts of interest, permitted and prohibited acts and language, review and withdrawal conditions, and final accountable owner. Without a record there is no decision—only an event for which a story may later be invented.

The judgment of this chapter is:

Authority may be delegated. Accountability may not. Automation can perform operations; judgment and the duty to answer for them remain human.

Governance refuses some work, narrows some claims, declines some clients and withdraws approval where necessary. This is not weakness. It is the moment at which authority does its work.

No decision, however, remains valid forever. Evidence ages, protocols become obsolete, interventions create maintenance debt, and what is true today may cease to be current tomorrow. Time is therefore the outer ring.