Manifold Grace

RESEARCH & KNOWLEDGE · AUTHOR-ORIGINATED CONCEPT RECORD

Continuous Agent-Assisted AI Architecture Assurance and Controlled Improvement

Concept conceived before 2026 (author-reported; exact date not yet independently established). Public research record created: 11 November 2026. This record presents the researcher's own proposal, articulated in discussion before this publication, for using specialised AI agents to inspect an AI system's actual architecture, assess quality and compliance, report verified findings in a structured system-update format and support controlled improvements. This page documents an original articulation and proposed synthesis, not a claim that every underlying technique is historically unprecedented or that a complete implementation exists.

Central research question. Can independent auditing agents maintain an accurate, tamper-evident representation of an AI system's architecture, quality and legal compliance, detect undisclosed or hidden discrepancies, and transfer verified findings into a controlled improvement process without permitting the audited system to disable or modify its oversight?

Proposed architecture

AI system under examination
Model · infrastructure · deployment · components · dependencies · tools · data flows
↓ Authorised, independently verified evidence
Specialised audit agents
Architecture discovery · quality evaluation · security inspection · compliance mapping · anomaly detection · provenance
↓ Evidence-bounded findings
Structured, signed quality and architecture report
System/version identity · actual vs declared architecture · test results · applicable controls · evidence gaps · uncertainty
↓ Independent controls and approval
Externally enforced rules and authorised remediation
Versioned policies · separated duties · protected logs · legal review · correction verification
↓ Approved update and re-evaluation
Controlled system improvement
Configuration, retrieval, tests, software, or separately approved fine-tuning/training changes

Author's proposed contribution

The proposed integration is a continuous feedback loop between independent architecture-aware assurance and controlled system quality improvement. Its distinctive investigable emphasis is on whether agents can discover what has actually been deployed, identify discrepancies with declared architecture and binding requirements, report those discrepancies in a verifiable machine-readable record, and enable a system to improve through a separate, authorised change pipeline.

The architecture distinguishes observation, assessment, enforcement, authorisation, modification and follow-up verification. It requires evidence from authorised inventories, configuration, attestation, logs and testing; conversational self-description alone is insufficient. Some internal components may remain unobservable, and an auditor must explicitly report those limits.

Governance and implementation boundaries

RequirementDesign principleLimitation
Legislation and mandatory guardrailsSigned, externally enforced policy with authorised, traceable updates and separation of dutiesNo policy is literally immutable against every privileged attacker; applicable laws may change.
Architectural truthReconcile declared architecture with attestations, software inventories, deployment manifests and runtime evidenceIncomplete access can produce uncertain findings.
Agent independenceAuditors lack unilateral power to alter controls or deploy model changesAuditing agents must themselves be tested and supervised.
Self-improvementTranslate reports into proposed corrections, then run tests and require approval before releaseReceiving a report does not by itself retrain model weights.
Privacy and safetyLeast privilege, protected logs, limited disclosure and redressAccountability must not become unrestricted surveillance.

Prior work and overlap audit — initial, non-exhaustive

The references below document substantial prior art. They prevent overclaiming and identify a more precise question for original experimental contribution.

SourceFirst publicationEstablished overlapRemaining question for this proposal
NIST AI Risk Management Framework2023AI risk governance, measurement and managementArchitecture-aware continuous multi-agent verification
ISO/IEC 420012023AI management-system requirementsEvidence of deployed-system behaviour, not just management processes
Anthropic Petri2025Automated model-behaviour auditingInspection of concealed architecture and controlled update handoff
WEF agent authorisation playbookMay 2026Auditable agent authorisation profilesVerification against actual system architecture
MITRE Continuous Remote AttestationJuly 2026Hardware-rooted, ongoing workload verificationConnection to model-quality tests and controlled learning
IETF Agent Auditing Architecture draft2026Distributed agent audit records, delegation and attestationCross-layer architecture reconciliation and remediation tests
From Assistants to Auditors (GRAAF)2026Multi-source agent-assisted audits and evidentiary reportingFalsifiable evaluation of quality-improvement feedback under strict separation of powers
LATTICE governance-first architecture2026Deterministic controls, policy enforcement and audit trailsDetection of hidden configuration mismatch and independent update assurance

Proposed falsification tests

Candidate experiments would evaluate detection of mismatches between declared and actual architecture; identification of unauthorised policy changes; accuracy under missing or misleading evidence; audit false-positive and false-negative rates; repeatability across model families; signed report validity; and the effectiveness of independent review gates. No experimental results are claimed here.

Publication and provenance status

Status: public research proposal and initial related-work review. The proposal belongs to the author of Manifold Grace. Assistance with synthesis, editing or publication does not establish independent AI authorship of the originating suggestion. The author reports having developed the underlying idea before 2026, prior to this public write-up, while other academic and institutional matters delayed further work on it. Git history records publication and subsequent amendments, not the original conception date. Earlier dated evidence, patentability and distinct technical novelty require separate verification.

Related comparative AI assurance analysis · Research directory · Productivity log