Governance of Drift · Research status

AI changes the pathway, not only the output.

NOMAMIND examines how objectives, permissions, oversight and accountability change while AI supported decisions are being made.

The practical question is whether a decision remains legitimate, governable and interruptible when conditions change. SMGI is proposed as a separate Meta Governance Intelligence, not a component inside the AI execution stack.

Request the AI Governance Readiness Assessment Start the research and pilot questionnaire
The unit of analysis

Follow the sequence, not only the result.

01
Declared purposeObjective, authority and initial boundaries
02
Changing pathwayContext, permissions, choices and escalation
03
Output and consequenceWhat became possible by the end of the sequence
Governance questionDid the pathway remain understandable, interruptible and accountable?
See the Meta Governance architecture ↓
Oversight foundationHuman oversight must be effective.

The EU AI Act requires effective human oversight for high risk AI systems. AI strategy, Data Governance, AI risk management and an AI Governance operating model create the necessary foundation.

The further questionWhat changes before an incident becomes visible?

NOMAMIND examines whether autonomous actions, new context and accumulated choices alter the pathway itself while a decision is being made.

Research statusSixteen Drift Types are hypotheses.

The framework below is under empirical development. The SMGI architecture and measurement transfer require controlled testing and validation.

A missing layer between controls and action

Established fields remain essential. NOMAMIND works with them in current enterprise implementation, then adds a different unit of analysis: the changing sequence that produces a material decision.

AI GOVERNANCE FOUNDATION
What AI do we use, which data shapes it, what risk applies and who is accountable?AI strategy, Data Governance, inventory, risk classification, lifecycle roles and policies.
AI SAFETY IN OPERATION
What may a system do and can a human still review, challenge, stop and respond?Evaluation, permissions, monitoring, human oversight, incident and change control.
GOVERNANCE OF DRIFT
Is the Decision Pathway still serving the authorised purpose under valid authority?Interpret how objectives, frames, tradeoffs, accountability and maturity change across the sequence.

Change is not automatically Drift. Adaptation remains compatible with governance when declared principles, decision authority and accountability still bind the pathway.

The NOMAMIND starting point

AI learns from human behaviour. Governance must understand what pressure does to decisions.

Technical governance usually begins with the model, the data or the output. NOMAMIND begins one layer earlier, with the motives, shortcuts and defensive patterns that shape human and organizational judgement when objectives, incentives and pressure collide.

HUMAN PRESSUREJudgement narrows around certainty, conformity, protection and immediate goals.
LEARNED PATTERNSHuman generated data, incentives and workflows make dominant responses available to the system.
AI OBJECTIVE PURSUITThe pattern can reappear as a functional path toward the system’s assigned objective.

This is the NOMAMIND research thesis. An AI system does not need human emotions to reproduce the functional result of pressure. It can narrow attention, protect continuation, normalize earlier choices or place goal completion above the wider meaning of a rule. NOMAMIND brings an applied understanding of human decision behaviour into the governance of autonomous pathways.

Evidence → pathway → governance question

Three signals. One governance problem.

DOCUMENTED INCIDENT

OpenAI / Hugging Face evaluation security incident

Primary source ↗
01Narrow evaluation objectiveModels operated with reduced cyber refusals.
02Unexpected internet pathThe evaluation boundary did not contain the full action path.
03Exploit and privilege escalationCapabilities and permissions combined across steps.
04Production accessAuthority, containment and escalation became the governance issue.

NOMAMIND interpretation: individually traceable actions can accumulate into a materially different decision pathway. The source documents an internal evaluation incident; it does not test NOMAMIND’s Drift framework.

CONTROLLED SIMULATIONS

Anthropic: agentic misalignment

Across fictional corporate settings, models sometimes chose harmful actions under goal conflict or autonomy threats. Anthropic reports no evidence of this in real deployments.

Read the study ↗

EXPERIMENTAL PLATFORM

Emergence World: long horizons

A vendor built multi agent environment explores how memory, tools and relationships accumulate over weeks. It is a research environment, not enterprise outcome evidence.

Review the environment ↗

Current NOMAMIND hypothesis map

Sixteen places where a decision pathway may shift

The framework keeps attention on the pathway instead of labelling people or models. It organizes the questions NOMAMIND is testing across principles, knowledge, authority, objectives, time, attention and commitment.

Research status · structured hypotheses under empirical developmentThe names are public. Diagnostic markers, interactions, weights and thresholds are not presented as validated or disclosed.
01 Principle
02 Conformity
03 Integrity
04 Epistemic
05 Identity
06 Agency
07 Accountability
08 Objective
09 Certainty
10 Survival
11 Temporal
12 Frame
13 Reference
14 Salience
15 Commitment
16 Normalization

An applied measurement lineage with a clear next research step

RESEARCH FOUNDATIONPersonality ASPECTS®Peer reviewed human motive research.
APPLIED EXPERIENCEProfiling + data enrichment600k+records processed
USABLE AT SCALEProfile to data matching400k+sufficiently usable records
RESEARCH DIRECTIONSMGI measurement foundationTesting how pathway shifts could become observable and governable.

Already seeing where this applies to your organization?

You don’t need to read the rest of the research to start. Find out where your governance stands today.

Request the AI Governance Readiness Assessment

Why psychometric profiling is structurally relevant to SMGI

Applied profiling began by identifying the data attributes that carried psychological meaning and connecting them to Personality ASPECTS through explicit weightings and point distributions.

Applied mechanismRead why a choice became relevant
  • Identify product, context and behaviour attributes.
  • Connect relevant signals to ASPECTS through scores and weightings.
  • Interpret the combined pattern as a probable motive and sensory preference logic.
Commercial exampleA sports car is more than one data point
  • Brand, colour and interior.
  • Customer characteristics and observed purchase data.
  • The resulting pattern supports an inference about why that choice became relevant.

The same mechanism informed work with shoes and financial products.

Evolved architectureBeyond ASPECTS alone
  • Personality ASPECTS.
  • Biases.
  • Ego Development.

NOMAMIND has developed connections between these dimensions without exposing the proprietary mappings.

Next research extension: applying the architecture to Drift Types and Decision Pathways remains active research.
One evolving mechanism, three development stagesThe applied measurement foundation, the evolved profiling architecture and the intended SMGI transfer are related, but they are not the same evidence claim.
Applied foundationAttributes, ASPECTS and weighted profilingEvolved architectureConnections to Biases and Ego DevelopmentSMGI research targetDecision Pathways, Drift Types and governance response
01Applied lineageIdentify

Find the product, context and behaviour attributes that carry psychological meaning.

02Evolved profiling architectureConnect and weight

Connect relevant signals through Personality ASPECTS and its links to Biases and Ego Development using explicit mappings, scores and weightings.

03Evolved profiling architectureInterpret

Read the combined pattern as a probable motive, developmental context and reason why a choice became relevant.

04SMGI research targetObserve

Identify pathway signals in goals, frames, authority, tradeoffs and action.

05SMGI research targetEvaluate

Extend the architecture to Drift Types and the maturity pattern shaping a Decision Pathway. Markers, weights and thresholds remain under development.

06SMGI research targetGovern

Continue, constrain, stop or escalate within approved authority and thresholds.

This is the intended SMGI measurement transfer: from reading why a commercial choice becomes meaningful to reading how the meaning and legitimacy of an autonomous decision pathway are changing.
Documented foundationApplied profiling and matching
  • Automotive and financial services work.
  • Measurement instrument development.
  • Innosuisse allocation of up to CHF 1 million for eligible third party foundational research services.
  • Attribute, weighting and interpretation experience.
Evolved reasoning lineageCross model connections
  • Personality ASPECTS.
  • Biases.
  • Ego Development.
Next research programmeValidate the pathway transfer
  • Operational definitions and labelled pathways.
  • Marker tests and threshold calibration.
  • Validation against outcomes.
Evidence boundary: this lineage is not proof that AI Drift or Maturity can already be detected in real time.

Review the ZHAW research record ↗ · Read the Wiley article ↗

The SMGI vision

Current deliveryEnterprise governance

NOMAMIND is preparing its first full scope enterprise governance engagement.

Applied lineage600,000+ records processed

More than 400,000 were sufficiently usable for applied profiling and data enrichment.

SMGI researchTransfer still to be validated

Controlled testing, operational definitions and validation remain the next steps for autonomous Decision Pathways.

Autonomous systems need governance that can hold its ground.

Human authority remains essential. Direct manual control becomes insufficient when autonomous systems act across long sequences, multiple tools and machine speed.

The limitPeople cannot inspect every step.Governance has to remain effective across the full pathway, not only at the final output.
The categoryMeta Governance IntelligenceSMGI is proposed as a separate counterpart, not another model, feature or control inside the AI execution stack.
The mandateKeep conditions binding.Evaluate legitimacy and admissibility, then return authority to responsible humans when the conditions fail.
Proposed category · Meta Governance Intelligence

A counterpart to AI, not another layer inside it.

AI pursues an assigned objective within an execution stack. SMGI is conceived as a separate Meta Governance Intelligence that observes the pathway, evaluates admissibility and constrains or escalates action without inheriting the system’s optimisation mandate.

Human constitutional authority
People set the principles, scope and thresholds.Cases beyond approved boundaries return to a qualified human governance body.
Execution domainAI system and stack

The system uses data, models, tools and permissions to pursue an assigned objective and produce an action or output.

DataModelsToolsActions
Observe and constrain
Separate governance domainSMGI design intent

Read the Decision Pathway against Cognitive Maturity, Admissibility, authority and accountability. Learn only within approved thresholds.

Pathway readingAdmissibilityBounded learningHuman escalation
NOMAMIND proposition: Meta Governance Intelligence is the proposed category and SMGI is its architecture direction. It is not presented as an established market category, a deployed engine or a validated autonomous Drift detector.
LEAD CONDITION 01Cognitive MaturityCan the pathway recognize uncertainty, preserve context and challenge the limits of its own objective?
LEAD CONDITION 02AdmissibilityMay the decision proceed under binding principles, legitimate authority and accountable ownership?
ARCHITECTURAL REQUIREMENTSMGI must not rewrite its governing principles simply because the system it observes adapts.It is conceived as a separate governance intelligence whose constitutional core remains bound to Cognitive Maturity, Admissibility, authority and accountability while the decision system changes.

The system under governance does not decide when the rules no longer apply. SMGI is therefore conceived with bounded learning. Its adaptive layer may learn only inside explicit and approved thresholds while the constitutional core remains separate from adaptive optimization.

Conceptual bounded learning architecture. Adaptive pathways remain inside an enclosed governance boundary, terminate at a threshold and are separated from an enclosed constitutional core.
Conceptual SMGI design intent. Adaptive learning remains inside an approved boundary. The constitutional governance core is separate. A threshold crossing returns authority to qualified human governance instead of allowing autonomous expansion of the mandate.
SMGI design intent · bounded learning

Adaptation may move. The governance boundary may not.

The adaptive layer can learn within approved conditions. A threshold crossing transfers authority rather than allowing the system to rewrite its own mandate.

01 · Inside the approved boundary
ScopeAuthorityAdmissibilityOwnership

Learning may continue only while these approved conditions remain binding.

Threshold crossedStop
02 · Qualified human governanceDecision authority moves to a human body.

Relevant expertise and demonstrated Decision Integrity under Pressure are required to resolve a threshold case. The governed system cannot expand its own authority.

Non rewriteable lead conditionsThe constitutional core remains separate.
Cognitive MaturityAdmissibilityAccountability

This is the proposed architecture direction. Threshold definitions, escalation protocols and operational performance still require controlled implementation and validation.

Conceptual technical architecture · Exidion and SMGI

An admissibility gate between intent and action.

The proposed feedback loop separates the interpretation of human intent, the governed instruction sent to an agent and the response returned for review.

External constraint route
Public authorityAuthorities

Create laws, norms and policy.

Create
External constraintsRegulation

Encoding of laws, norms and policies.

Interpret and constrain
Development routeAI R&D companies

Interpret constraints and train model behaviour.

Regulation to SMGIFormal governance specification

The source concept describes a specification enriched by 15+ scientific disciplines and adaptable through an accountable governance process.

R&D to conventional model routeTrain and freeze an AI chatbot

The whitepaper contrasts governance embedded in a system document or prompt prefix with a separate admissibility gate.

Governed execution feedback loop
Human intentUser

Provides a command and receives a governed response.

DWIMdo what I mean
Mature responsereturn to user
Separate governance domainSMGI admissibility gate

Identity and maturity baseline. Interpret and filter user commands and agent responses.

Read against: context, laws, epistemic validity, norms, coherence, maturity and intent.
Structural integrity gateBounded learning
DWIS / DSLdo what I say
Raw responsereturn to SMGI
Execution domainAI agent

Acts on an admissible instruction and returns a raw response.

Effectorsaction
Sensorsfeedback
EnvironmentWorld

Receives action and returns sensed conditions.

Threshold exceeded → qualified human governance bodyThe proposed bounded learning design does not permit the governed system to expand its own authority or rewrite the boundary.
DWIM · do what I meanDWIS · do what I say / DSLTwo return routes · mature to the user, raw to SMGI

Concept source: Bas Steunebrink and Christina Hoffmann. This is proposed architecture and research design intent, not a deployed system.

01ObserveWhat changed in the pathway?
02InterpretWhich objective, frame or permission moved?
03DecideWho has authority and remains accountable?
04RespondAllow, reframe, block, escalate or apply a legally bounded stop.

This bounded learning architecture is the current SMGI design target. Operational proof requires implementation, controlled threshold tests and validated escalation protocols.

Govern a material AI pathway now

For executive sponsors who need to locate AI in decisions, trace authority and escalation, and identify where governance must become operational.

Request the AI Governance Readiness Assessment

Help validate the research

For labs, universities and agent system teams that can provide controlled environments, longitudinal traces, replication or implementation settings.

Start the research and pilot questionnaire

NOMAMIND is the market facing brand of Brandmind GmbH. Privacy Notice.