Manifold Grace

RESEARCH & KNOWLEDGE · HISTORICAL EVIDENCE RECORD

AI Auditing: Historical Development, Architectural Assurance and Original Contributions

Record prepared 11 November 2026. This research page distinguishes dated evidence of earlier AI-auditing interests from later documentation of a particular agent-assisted architecture. The author reports conceiving the underlying idea before 2026. The dates below identify source records or publications, not necessarily original conception.

Primary methodological distinction: (1) demonstrated historical discussion, (2) author-reported contribution, (3) independently verified technical implementation, and (4) novel technical contribution are different propositions. A dated assistant summary of a subject does not by itself prove the author's specific earlier technical design.

Historical development: source-bounded chronology

PeriodResearch development or source leadEvidence status and limits
27 November 2024AI Bias and Solutions: model bias auditing, fairness metrics and explainability; Trusted Data Sources: algorithm auditing mentioned.Dated ChatGPT history search results and screenshot excerpts supplied by the author. Visible text supports discussion of the subjects; complete conversations and the author's initiating questions were not independently examined.
December 2024People vs Profit Businesses (6 December): independent audit firms; Housing Rights and Legislation (12 December): dataset auditing.Dated search-result leads provided in screenshots, not verified full transcripts. Additional historical entries also appeared but are not used here to infer architectural details.
26–31 January 2025Cultural Values and Gender (26 January): auditing and enforcement; MPhil PhD Application Guide (31 January): academic context for AI auditing.Search-result titles, dates and short snippets visible in screenshots; full text not independently retrieved.
2024–2025 academic contextMA Week 8 material addressed algorithmic bias, fairness measures, explainability, auditing and technical limitations.Author-provided account; relevant course document and its precise date not examined for this public record.
11 November 2026Continuous Agent-Assisted AI Architecture Assurance and Controlled Improvement formally documented.Publicly documented architecture and associated GitHub history; this is not evidence of first conception in 2026.

What the proposal investigates

Specialised agents would inspect an AI deployment's architecture and operation through authorised evidence; reconcile the observed design against declarations, policies and compliance duties; detect discrepancies or hidden violations where access permits; produce signed, structured assurance findings; and pass verified corrections into a separately authorised improvement and testing pipeline. The research question is whether this combination can improve evidence completeness, reliability and independence in complex AI assurance.

The design prohibits ordinary audit agents from unilaterally rewriting binding controls or deploying updates. No system can be presumed entirely inspectable or immune to privileged alteration; missing access, false positives, uncertain causal conclusions and maliciously inaccurate declarations remain central test conditions.

Historical intellectual activity candidates (2024–2025)

This register transcribes and organises sixteen candidate areas of intellectual activity identified in a later conversation. They are not sixteen verified original inventions. Numbers 8–9 belong to December 2024 despite being included in the original twelve-item list; numbers 13–14 expand those same December themes and are not independent evidence of separate inventions. The screenshot evidence establishes that relevant topics were discussed, not who independently introduced each idea.

No.Candidate intellectual workHistorical evidence leadAttribution / verification status
1Algorithmic bias identificationMA Week 8 learning outcomes / seminar promptsEstablished subject studied; independent contribution not verified
2Multi-layer bias analysis (data, categories, model, deployment)Week 8 discussionsAnalytical framework candidate; authorship of formulation unverified
3Continuous algorithmic auditingAI Bias and Solutions — 27 Nov 2024Dated subject discussion; own proposal needs transcript verification
4Disparate-impact monitoringAI Bias and Solutions — 27 Nov 2024Established technique; potential contextual application
5Fairness measurement and competing definitionsMA seminar prompt on inadequacy of remediesPotential critical analysis; own argument not verified
6Explainability and accountabilityAI Bias and Solutions — 27 Nov 2024Potential philosophical argument; authorship unverified
7Data reliability and provenanceTrusted Data Sources — 27 Nov 2024Potential research direction; precise contribution unverified
8Independent AI auditing organisationsPeople vs Profit Businesses — 6 Dec 2024December lead included in initial 12; not a November-dated original claim
9Dataset auditing and institutional accountabilityHousing Rights and Legislation — 12 Dec 2024December lead included in initial 12; not a November-dated original claim
10Algorithmic injustice and access to opportunityMA coursework / Herzog readingApplied ethical investigation candidate
11Language, classification and cultural biasMA coursework / Crawford readingPhilosophical investigation candidate
12Limits of technical bias remediesMA Week 8 outcome and seminar questionPotential critical synthesis; precise argument unverified
13Independent institutional oversight of AI developersPeople vs Profit Businesses — 6 Dec 2024December extension; overlaps candidate 8
14Algorithmic auditing in housing and public decision-makingHousing Rights and Legislation — 12 Dec 2024December application; overlaps candidate 9
15Cultural values, fairness and enforceable accountabilityCultural Values and Gender — 26 Jan 2025January 2025 extension; original formulation unverified
16Formal academic methodology for AI auditingMPhil PhD Application Guide — 31 Jan 2025January 2025 extension; original formulation unverified

Archival designation: AI Ethics, Algorithmic Bias and Independent Auditing — Intellectual Work, November 2024, with separately dated December 2024 and January 2025 developments. Source basis: author-supplied excerpts and screenshots of ChatGPT search results, plus author-described MA Week 8 materials. The original complete transcripts and course handouts have not been independently retrieved for this entry. The author's own initiating questions, substantive arguments, proposed mechanisms and precise chronology require confirmation before originality or priority is claimed.

Interpretive rule: Studying a pre-existing concept, critically extending it and independently proposing an implementable mechanism are three separate contribution types. The 2026 publication documents the later architecture; it does not retrospectively assign every 2026 detail to 2024.

Initial prior-art comparison

Published workYearEstablished contributionGap requiring testing, not presumed novelty
NIST AI Risk Management Framework2023Structured AI risk identification, evaluation and management.Cross-layer live architecture verification and automated, independently authorised remediation.
ISO/IEC 420012023AI management-system controls and assurance processes.Proof of compliance at runtime for opaque or heterogeneous architectures.
NIST Generative AI Profile2024Generative-AI-specific risk management.Directly measured accuracy of independent agent-produced architecture and compliance reports.
Building and evaluating alignment auditing agents2025Autonomous auditing agents and experimentally measured strengths and limitations.Verifiable inspection of deployed configuration, non-model components and subsequent correction controls.
Anthropic Petri2025Repeatable multi-turn behavioural probes by automated agents.Integration with architectural attestation and compliance-enforced improvement.
Continuous Remote Attestation2026Verification of computing workload and infrastructure integrity.Linking integrity evidence to nuanced model-quality assessment and validated learning updates.
Architecture for auditing agent delegation and interactions2026Traceable delegated-agent activity and audit architecture (draft).Evidence-backed reconciliation of claimed and observed AI architectures, linked to constrained remediation.

Testable contribution and falsification criteria

A systematic overlap review should ask what published systems already accomplish, then specify a remaining limitation: Can independent agents detect undeclared architecture changes or breaches of externally enforced controls, preserve provenance under partial access and improve quality without compromising oversight? Candidate metrics include discrepancy recall and precision, missing-evidence reporting, audit reproducibility, false-positive rates, remediation approval integrity and quality change after redeployment. These are proposed tests; no results are asserted.

Evidence integrity and attribution notes

Historical 2024–2025 chat references come from author-supplied screenshots showing titles, search dates and fragments. They are useful leads but not linked public transcripts; avoid deriving exact original wording, chronology of invention or technical features not shown. A full original conversation or dated independent artefact could strengthen the provenance record. The author's reported pre-2026 conception is preserved without changing the date of publication. Publication is not a patent-priority determination or proof of global novelty.

Technical architecture proposal · Comparative AI quality study · Research directory · Productivity log entry