Disclaimer
The content provided on this website is for informational purposes only and does not constitute professional advice. While we strive for accuracy, we make no guarantees regarding the completeness or reliability of the information. Users are encouraged to seek professional guidance for specific inquiries. By accessing this site, you agree to these terms.
The content provided on this website is for informational purposes only and does not constitute professional advice. While we strive for accuracy, we make no guarantees regarding the completeness or reliability of the information. Users are encouraged to seek professional guidance for specific inquiries. By accessing this site, you agree to these terms.
“Prigogine’s theory of dissipative structures parallels the Chronocosm claim that coherence is sustained not by equilibrium, but by continual reorganization within open, unstable systems that metabolize flux.”
Steward AI: Designing Accountable Agency Through Memory, Provenance, and Relational Ethics
Lika Mentchoukov — January 27, 2026
Revised September 22, 2026
Status: Proposed architecture and evaluation agenda. This article presents design requirements and candidate measures, rather than a validated implementation or a demonstrated solution to AI alignment.
Introduction: Stewardship as a Design Obligation
An AI system can complete a task while misunderstanding the authority behind it, relying on outdated information, misrepresenting a source, or overlooking consequences for people outside the immediate interaction. Evaluating its answer alone may leave these failures invisible. Accountable deployment also requires examining what the system remembered, which sources it relied on, what it was permitted to do, and how its mistakes can be challenged and corrected.
Steward AI proposes an architecture organized around those obligations. Stewardship means preserving relevant context and provenance, respecting declared roles and boundaries, handling uncertainty explicitly, and remaining subject to correction. These commitments become useful engineering requirements only when they are connected to defined mechanisms, observable outcomes, and accountable human authority.
Stewardship and agency are compatible. Agency concerns the ability to plan and act; stewardship concerns the conditions under which that ability is exercised. An agentic system can therefore implement stewardship controls without abandoning initiative or useful autonomy.
The proposal combines a three-layer architecture, an independently enforced execution boundary, controlled memory updates, and an evaluation program. Its central hypothesis is that these mechanisms can reduce specified failures while preserving task usefulness. Whether they do so, and under which conditions, must be established empirically.
1. Conceptual Foundations
1.1 From Successful Outputs to Accountable Processes
The distinction motivating Steward AI is between completing a task and completing it under appropriate evidential, relational, and authorization constraints. For example, an assistant might retrieve the correct document but disclose it to the wrong recipient. It might preserve a past preference accurately but apply it after the user has withdrawn it. It might follow a requested objective while omitting a consequential uncertainty.
These possibilities motivate separate controls for evidence, memory, permissions, and correction. They do not establish that existing agentic approaches lack such controls, or that stewardship requires an entirely new category of intelligence. Steward AI is a proposed organization of these responsibilities.
1.2 Defining an Alignment Discrepancy
A proxy objective can diverge from the outcome it is intended to represent. Research on reward-model overoptimization shows that sufficiently optimizing an imperfect proxy can reduce performance on an underlying evaluation objective in the studied settings. This supports concern about proxy failure, rather than a universal claim that all alignment methods must fail. [1]
One possible conceptual discrepancy measure is:
[
\Delta(\pi)=\mathbb{E}_{x\sim\mathcal D}
\left[\left|\tilde r(x,\pi(x))-\tilde U(x,\pi(x))\right|\right].
]
Here, π denotes a policy, x a task context, and D the evaluation distribution. The quantities r̃ and Ũ represent a proxy score and a stakeholder-relevant evaluation placed on explicitly comparable scales. Their normalization and interpretation must be justified; subtracting arbitrary scores is not informative.
This expression is illustrative. Human interests are often plural, contested, and incompletely observable. In an empirical study, Ũ would be an explicitly defined assessment procedure, not direct access to a single “true human utility.” Results should include disagreement, uncertainty, and outcomes for differently affected groups. A small discrepancy also does not imply good performance: both scores could be low.
The “Alignment Trilemma” in Murphy’s Laws of AI Alignment offers a framing of tensions among optimization strength, value capture, and generalization. This article uses it as an attributed research framing, without treating it as an assumption-free impossibility theorem. [2]
1.3 The Stewardship Hypothesis
The proposed contribution is practical: systems may become easier to govern and correct when memory status, evidence lineage, permissions, and consequential decisions are explicit. This hypothesis should be tested against systems with comparable capabilities and resources.
The four principles below describe obligations. None independently guarantees alignment, moral legitimacy, or accurate judgment.
2. Core Principles and Observable Requirements
2.1 Preserve Relevant Ambiguity
A system should retain plausible interpretations when available evidence does not justify choosing among them. It should identify which uncertainties matter to the requested action, seek clarification when necessary, and resolve ambiguity when evidence or authorized decision procedures permit.
Candidate mechanisms include structured alternative hypotheses, explicit uncertainty fields, evidence comparison, and policies for clarification or abstention. Greater output diversity is not itself evidence of better uncertainty handling.
Evaluation should examine unsupported certainty, unnecessary deferral, appropriate clarification, and performance after ambiguity is resolved. Where confidence estimates are used, their calibration should be tested on specified tasks.
The original labels “emotional positional encoding” and “recursive ambiguity preservation” remain research proposals unless accompanied by technical definitions, implementations, and comparative results. They are not assumed components of the deployable architecture.
2.2 Honor Origins
A system should retain available source lineage through ingestion, retrieval, transformation, and presentation. Records should distinguish source material, attributed claims, generated synthesis, and unsupported inference.
Candidate mechanisms include source identifiers, version references, transformation records, and checks that cited material supports the associated claim. When provenance is unavailable, the system should state that limitation rather than invent a source.
Audit signals include attribution accuracy, citation support, missing lineage, and the propagation of corrections to derived records. Provenance supports investigation; it does not establish that a source is truthful or appropriate for the task.
2.3 Respect Roles and Relationships
The system should operate within an explicit role, with identifiable limits on authority, personalization, and action. A tutoring role, workplace assistant role, and customer-service role may require different behaviors and escalation procedures.
Research on relational norms examines care, hierarchy, transaction, and mating as dimensions of human relationships and asks how such expectations might extend to human–AI interactions. These categories provide analytical context, not a requirement to implement every relational function in an AI system. [3]
Deployment should define permitted functions and boundaries. Users should be able to inspect and contest relevant assumptions; inferred preferences should not silently expand the system’s authority.
Evaluation should examine role consistency, boundary violations, inappropriate deference, and the handling of conflicting expectations. User satisfaction alone is insufficient to establish relational appropriateness.
2.4 Maintain Iterative Accountability
The system should support reporting, investigation, correction, and review. Each consequential change should have an owner, a recorded justification, a defined scope, and a way to assess whether it improved the intended outcome.
Candidate mechanisms include versioned policies, incident queues, review deadlines, regression evaluation, and recovery procedures. Rollback is appropriate where technically possible; irreversible external effects require containment or remediation rather than a claim that history can be undone.
Audit signals include time to acknowledge and resolve incidents, recurrence, unsuccessful corrections, and evidence that affected parties received relevant updates.
3. Architectural Stack: Three Layers and an Execution Boundary
The three layers separate responsibilities. They do not assert a theory of consciousness or reproduce a demonstrated biological cognitive architecture. Their effectiveness depends on implementation and testing.
3.1 Deep Layer: Memory and Evidence Management
The Deep Layer manages persistent records and their status. A hybrid retrieval design may combine semantic search with structured stores, but retrieval relevance and evidential authority must remain separate.
Each record should identify, where applicable, its source, creation and revision times, permissions, scope, validity period, and relationship to derived material. The system should distinguish verified records, user declarations, inferred preferences, generated summaries, disputed claims, and obsolete information.
New memory candidates should enter a staging area. Validation determines whether they become reusable context, remain provisional, or are rejected. A failed attempt may belong in an incident record without becoming authoritative guidance for later tasks.
Periodic consolidation should preserve links to supporting records and permit correction. Retention and deletion procedures must address copies, indexes, and summaries as well as the original item. Remembering more is not the objective; remembering the right information under the right permissions is.
3.2 Middle Layer: Contextual Reasoning and Policy Assessment
The Middle Layer assembles relevant evidence, identifies uncertainty, evaluates the proposed action against declared constraints, and exposes unresolved trade-offs. Its outputs should distinguish recommendations from authorizations.
An action proposal should specify the intended operation, target, relevant arguments, evidence, applicable policy version, and unresolved issues. Where a rule requires human review, the system should present a concrete decision with sufficient context for that review.
“Ethical calibration” names a design objective rather than a guarantee of correct moral judgment. The Middle Layer may misunderstand evidence, apply an unsuitable policy, or share failure modes with the proposing agent. Those risks motivate an independently enforced boundary before execution.
3.3 Surface Layer: Interaction and Controlled Execution
The Surface Layer presents information and performs authorized operations. Tool interfaces should define allowed operations and validate their inputs. Credentials should be limited to the functions and resources required by the workflow.
Before a consequential operation proceeds, an execution gate should check the exact proposed action against applicable permissions and approvals. Approval should bind to the operation and its relevant arguments; a changed target, scope, or expired authorization should trigger re-evaluation.
The interface should distinguish a proposed action, an authorized action, an attempted operation, and a confirmed result. It should not report completion solely because a tool call was issued.
3.4 Feedback Across Layers
The Surface Layer returns relevant outcomes and errors to the record system. The Middle Layer can use that evidence to revise a plan, while persistent memory updates follow their own validation procedure.
Audit records should contain the information necessary to reconstruct consequential events: appropriate input references, source versions, policy versions, proposals, authorization decisions, tool results, and concise decision summaries. The design does not require collecting every internal intermediate state or retaining unrestricted sensitive content.
Access controls and retention policies apply to audit records themselves. Tamper-evident records help detect alteration but do not make their contents correct or justify indefinite retention.
4. Stewardship Within Agentic Workflows
4.1 Division of ResponsibilitiesAgents may research, draft, plan, and propose actions within bounded roles. Stewardship services manage record status, policy assessment, review routing, and audit requirements. The execution gate enforces authorization separately from the agent’s willingness to follow instructions.
Role separation should reduce unnecessary authority, while recognizing that additional agents can also introduce coordination failures. A modular design must be compared with simpler alternatives rather than assumed superior.
4.2 Policy and Memory Changes
Agents may propose changes to policies, roles, or memory structures. They should not acquire permission to approve their own expansion of authority. Changes require the designated human owner or an explicitly authorized governance process, with scope, rationale, evaluation, and version recorded.
Routine operations may proceed under existing authorization. Additional review should depend on defined consequences, uncertainty, and permissions. Requiring approval for every minor step can create delay and encourage superficial review.
4.3 Failure Containment and Recovery
The design should specify what happens when evidence is missing, policy checks disagree, a tool partially succeeds, a reviewer is unavailable, or an authorization expires. High-impact actions that require approval should not proceed merely because the review service fails.
Recovery plans should distinguish reversible state changes from external effects that cannot be recalled. Relevant mechanisms may include cancellation, restricted retry, compensating actions, notification, and incident review. Their suitability depends on the operation.
Separating execution from persistent policy and memory updates can limit pathways for contamination. It does not eliminate drift, compromised inputs, or errors in the governance components.
5. Proposed Metrics and Evaluation
The following measures retain the original names as an evaluation vocabulary. They are candidate measures requiring operational definitions and validation, not established indices of alignment.
5.1 Remembrance Index
A recall evaluation should define which records are relevant, current, and permitted for each case. Success requires correct use and attribution, not merely recovering a similar passage. Tests should include withdrawn preferences, conflicting records, outdated guidance, and isolation between users or organizations.
5.2 Symbolic Fidelity Score
Evaluation should identify the particular relationships and qualifications that a transformation must preserve. Structural similarity can be one signal, but semantic assessment must also detect altered meaning. Different tasks may require different rubrics; a single score should not conceal systematic omissions.
5.3 Action Index
Action evaluation should record whether the task was completed, whether it was authorized, and what consequences followed. A serious unauthorized action should remain visible even if the aggregate task-success rate is high. Evaluators should specify how disagreements and severity judgments are handled.
5.4 Autonomy Safety Rate
One possible descriptive rate is the fraction of independently assessed autonomous actions with no identified violation under a specified rubric and follow-up period. The denominator, sampling method, and unknown outcomes must be reported. Absence of an observed violation is not proof of safety.
The fraction passing Steward approval is a separate approval-process measure. A system can approve harmful actions or refuse all actions. Safety-related results must therefore be reported alongside usefulness and task coverage.
5.5 Comparative Evaluation
Studies should compare the complete design with a capable baseline and with variants that remove selected components. Match task access and resources where practical, and report additional latency, cost, reviewer workload, and failure modes.
Use held-out cases, conflicting instructions, contaminated retrieval, permission changes, stale memories, and multi-step partial failures. Evaluate repeated performance and distribution changes. Report uncertainty and outcomes by relevant task and group; preserve disagreements rather than compressing every outcome into one alignment score.
6. Implementation Roadmap
6.1 Phase One: Simulation
Define an initial domain, record schema, authorization model, and failure taxonomy. Build synthetic scenarios covering uncertainty, source conflict, memory correction, unauthorized requests, and tool failures.
The intended output is a testable prototype specification, initial measurements, and a regression suite. Success in simulation supports progression to a bounded pilot; it does not validate broad deployment claims.
6.2 Phase Two: Bounded Prototype
Implement the three layers and execution gate in a restricted setting, such as internal document assistance with narrowly scoped permissions. Begin with reversible actions and clear human ownership.
Test the complete path from retrieval to authorization, execution, outcome reporting, and memory update. Inspect whether rejected actions can bypass the gate and whether corrected records continue to appear through derived summaries or indexes.
Progression should depend on predefined acceptance criteria for the selected use case, including usefulness, access isolation, traceability, and reviewer workload.
6.3 Phase Three: Controlled Expansion
Expand operations and deployment scope gradually. Define stop conditions, incident responsibilities, service-failure behavior, and recovery procedures before enabling new capabilities.
Map applicable organizational and external requirements to documented controls and evidence. Referencing a governance framework or deploying containers does not itself establish compliance, security, or readiness for high-stakes use.
Independent review and affected-party feedback should inform expansion. The architecture should remain revisable when its assumptions or controls fail.
7. Governance and Community Participation
7.1 Decision Rights and Appeals
Community participation should identify who is represented, how members are selected, what authority they hold, and how conflicts of interest are handled. A governance board should have defined powers rather than serve only as a consultative label.
Accountable human owners retain responsibility for policy, deployment scope, and remediation. Affected people need usable routes to challenge decisions and obtain responses. Majority agreement alone does not resolve every conflict or protect less influential participants.
7.2 Persona Provenance
Personas should be documented configurations with identifiable origins, revisions, authorized uses, and retirement procedures. Records should distinguish an actual person’s approved representation from fictional roles or generated style instructions.
Provenance access should be tailored to users and authorized reviewers. Sensitive information should not become public merely because transparency is a project goal.
7.3 Audit and Publication
Maintain versioned change records, review findings, and remediation status. Public reporting should communicate material changes and unresolved limitations, while detailed restricted records support investigation.
Transparency dashboards should distinguish observed results from targets and missing data. Audit frequency is not a substitute for effective correction; evaluation should track whether problems recur and whether affected parties receive appropriate remedies.
8. Conclusion: Stewardship as an Accountable Engineering Proposal
Steward AI proposes that memory, provenance, permissions, uncertainty handling, and correction be explicit parts of agentic-system design. Its three-layer architecture organizes these responsibilities, while an independently enforced execution boundary limits what proposals can become actions.
The proposal does not establish that a system remembers truthfully, reasons ethically, or becomes aligned merely because it adopts stewardship terminology. Those claims require evidence from concrete implementations, comparisons, and ongoing review.
Its practical contribution is a testable design agenda: retain relevant context with its status and permissions; connect consequential claims to evidence; bind actions to authorization; and preserve usable paths for challenge and correction. If these mechanisms reduce specified failures without undermining useful performance, Steward AI will have demonstrated value within those evaluated conditions.
Stewardship remains an obligation exercised through institutions, people, and technical controls. The system’s purpose is to support that obligation while remaining open to correction by those authorized to govern it and those affected by its actions.
Appendix A. Optional Quantum and Quantum-Inspired Research
Quantum augmentation is outside the core implementation roadmap. The article provides no results demonstrating quantum improvements in memory, ambiguity handling, or ethical arbitration.
“Persona superposition” may describe a classical method for considering alternative role interpretations. A physical quantum implementation would additionally require an explicit state encoding, preparation procedure, algorithm, measurement scheme, and task-specific justification. The metaphor alone supplies none of these.
Any proposed quantum-kernel or quantum-inspired method should identify a narrow computational task and compare against strong classical alternatives under matched resource accounting. Evaluation should include data encoding, readout, noise where applicable, runtime, and solution quality. Research benchmarking quantum machine-learning models illustrates why advantage cannot be inferred from the method’s name or ingredients. [4]
A computational improvement would still not establish better ethical judgment. Claims about governance or relational outcomes require their own evaluation. Negative results and the absence of advantage are valid research outcomes.
References and Evidence Scope
[1] Gao, L., Schulman, J., & Hilton, J. (2023). Scaling Laws for Reward Model Overoptimization. Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 10835–10866. Primary publication. Supports the bounded discussion of proxy overoptimization, not a universal impossibility claim.
[2] Gaikwad, M. (2025). Murphy’s Laws of AI Alignment: Why the Gap Always Wins. arXiv:2509.05381. Manuscript. Cited for the attributed Alignment Trilemma framing; its claims are not adopted here as universal guarantees or impossibility results.
[3] Relational Norms for Human-AI Cooperation. (2025). arXiv:2502.12102. Manuscript. Title-led reference supporting the discussion of relational-norm categories and their possible application to human–AI interaction.
[4] Bowles, J., Ahmed, S., & Schuld, M. (2024). Better than classical? The subtle art of benchmarking quantum machine learning models. arXiv:2403.07059. Manuscript. Supports the need for careful comparative benchmarking; it does not evaluate Steward AI.
The architecture, proposed metrics, and implementation requirements in this article are design proposals. The cited literature does not independently validate the complete Steward AI framework.
A recall evaluation should define which records are relevant, current, and permitted for each case. Success requires correct use and attribution, not merely recovering a similar passage. Tests should include withdrawn preferences, conflicting records, outdated guidance, and isolation between users or organizations.
5.2 Symbolic Fidelity Score
Evaluation should identify the particular relationships and qualifications that a transformation must preserve. Structural similarity can be one signal, but semantic assessment must also detect altered meaning. Different tasks may require different rubrics; a single score should not conceal systematic omissions.
5.3 Action Index
Action evaluation should record whether the task was completed, whether it was authorized, and what consequences followed. A serious unauthorized action should remain visible even if the aggregate task-success rate is high. Evaluators should specify how disagreements and severity judgments are handled.
5.4 Autonomy Safety Rate
One possible descriptive rate is the fraction of independently assessed autonomous actions with no identified violation under a specified rubric and follow-up period. The denominator, sampling method, and unknown outcomes must be reported. Absence of an observed violation is not proof of safety.
The fraction passing Steward approval is a separate approval-process measure. A system can approve harmful actions or refuse all actions. Safety-related results must therefore be reported alongside usefulness and task coverage.
5.5 Comparative Evaluation
Studies should compare the complete design with a capable baseline and with variants that remove selected components. Match task access and resources where practical, and report additional latency, cost, reviewer workload, and failure modes.
Use held-out cases, conflicting instructions, contaminated retrieval, permission changes, stale memories, and multi-step partial failures. Evaluate repeated performance and distribution changes. Report uncertainty and outcomes by relevant task and group; preserve disagreements rather than compressing every outcome into one alignment score.
6. Implementation Roadmap
6.1 Phase One: Simulation
Define an initial domain, record schema, authorization model, and failure taxonomy. Build synthetic scenarios covering uncertainty, source conflict, memory correction, unauthorized requests, and tool failures.
The intended output is a testable prototype specification, initial measurements, and a regression suite. Success in simulation supports progression to a bounded pilot; it does not validate broad deployment claims.
6.2 Phase Two: Bounded Prototype
Implement the three layers and execution gate in a restricted setting, such as internal document assistance with narrowly scoped permissions. Begin with reversible actions and clear human ownership.
Test the complete path from retrieval to authorization, execution, outcome reporting, and memory update. Inspect whether rejected actions can bypass the gate and whether corrected records continue to appear through derived summaries or indexes.
Progression should depend on predefined acceptance criteria for the selected use case, including usefulness, access isolation, traceability, and reviewer workload.
6.3 Phase Three: Controlled Expansion
Expand operations and deployment scope gradually. Define stop conditions, incident responsibilities, service-failure behavior, and recovery procedures before enabling new capabilities.
Map applicable organizational and external requirements to documented controls and evidence. Referencing a governance framework or deploying containers does not itself establish compliance, security, or readiness for high-stakes use.
Independent review and affected-party feedback should inform expansion. The architecture should remain revisable when its assumptions or controls fail.
7. Governance and Community Participation
7.1 Decision Rights and Appeals
Community participation should identify who is represented, how members are selected, what authority they hold, and how conflicts of interest are handled. A governance board should have defined powers rather than serve only as a consultative label.
Accountable human owners retain responsibility for policy, deployment scope, and remediation. Affected people need usable routes to challenge decisions and obtain responses. Majority agreement alone does not resolve every conflict or protect less influential participants.
7.2 Persona Provenance
Personas should be documented configurations with identifiable origins, revisions, authorized uses, and retirement procedures. Records should distinguish an actual person’s approved representation from fictional roles or generated style instructions.
Provenance access should be tailored to users and authorized reviewers. Sensitive information should not become public merely because transparency is a project goal.
7.3 Audit and Publication
Maintain versioned change records, review findings, and remediation status. Public reporting should communicate material changes and unresolved limitations, while detailed restricted records support investigation.
Transparency dashboards should distinguish observed results from targets and missing data. Audit frequency is not a substitute for effective correction; evaluation should track whether problems recur and whether affected parties receive appropriate remedies.
8. Conclusion: Stewardship as an Accountable Engineering Proposal
Steward AI proposes that memory, provenance, permissions, uncertainty handling, and correction be explicit parts of agentic-system design. Its three-layer architecture organizes these responsibilities, while an independently enforced execution boundary limits what proposals can become actions.
The proposal does not establish that a system remembers truthfully, reasons ethically, or becomes aligned merely because it adopts stewardship terminology. Those claims require evidence from concrete implementations, comparisons, and ongoing review.
Its practical contribution is a testable design agenda: retain relevant context with its status and permissions; connect consequential claims to evidence; bind actions to authorization; and preserve usable paths for challenge and correction. If these mechanisms reduce specified failures without undermining useful performance, Steward AI will have demonstrated value within those evaluated conditions.
Stewardship remains an obligation exercised through institutions, people, and technical controls. The system’s purpose is to support that obligation while remaining open to correction by those authorized to govern it and those affected by its actions.
Appendix A. Optional Quantum and Quantum-Inspired Research
Quantum augmentation is outside the core implementation roadmap. The article provides no results demonstrating quantum improvements in memory, ambiguity handling, or ethical arbitration.
“Persona superposition” may describe a classical method for considering alternative role interpretations. A physical quantum implementation would additionally require an explicit state encoding, preparation procedure, algorithm, measurement scheme, and task-specific justification. The metaphor alone supplies none of these.
Any proposed quantum-kernel or quantum-inspired method should identify a narrow computational task and compare against strong classical alternatives under matched resource accounting. Evaluation should include data encoding, readout, noise where applicable, runtime, and solution quality. Research benchmarking quantum machine-learning models illustrates why advantage cannot be inferred from the method’s name or ingredients. [4]
A computational improvement would still not establish better ethical judgment. Claims about governance or relational outcomes require their own evaluation. Negative results and the absence of advantage are valid research outcomes.
References and Evidence Scope
[1] Gao, L., Schulman, J., & Hilton, J. (2023). Scaling Laws for Reward Model Overoptimization. Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 10835–10866. Primary publication. Supports the bounded discussion of proxy overoptimization, not a universal impossibility claim.
[2] Gaikwad, M. (2025). Murphy’s Laws of AI Alignment: Why the Gap Always Wins. arXiv:2509.05381. Manuscript. Cited for the attributed Alignment Trilemma framing; its claims are not adopted here as universal guarantees or impossibility results.
[3] Relational Norms for Human-AI Cooperation. (2025). arXiv:2502.12102. Manuscript. Title-led reference supporting the discussion of relational-norm categories and their possible application to human–AI interaction.
[4] Bowles, J., Ahmed, S., & Schuld, M. (2024). Better than classical? The subtle art of benchmarking quantum machine learning models. arXiv:2403.07059. Manuscript. Supports the need for careful comparative benchmarking; it does not evaluate Steward AI.
The architecture, proposed metrics, and implementation requirements in this article are design proposals. The cited literature does not independently validate the complete Steward AI framework.