Manifold Grace

RESEARCH & KNOWLEDGE / PROJECTS & ARCHITECTURE

Architecture, Assurance & Research Project Pathways

Working framework published 13 November 2026. A connected programme for studying software and AI architectures, comparing original source code with authorised educational derivatives, assessing legal and regulatory duties, testing cybersecurity and safeguarding controls, and documenting measurable outcomes. It supports academic research, professional learning, reproducible engineering and independent scrutiny. The project pathways below are proposals and reusable specifications unless explicitly linked to an implemented repository and tested result.

Existing project register · Research programme · Research questions · Dated change record.

One traceable chain from source to outcome

Evidence layerMinimum required recordWhy it matters
1. Published foundationOriginal paper, publication year, version, DOI or permanent URL, claims, limitations and related workPrevents undocumented borrowing and separates established findings from proposed contributions.
2. Upstream code and provenanceOfficial repository, immutable commit SHA or release, licence, copyright notices, dependencies, known issues, archived referenceEstablishes exactly which artefact was inspected or reused; public visibility does not automatically grant reuse rights.
3. Study and reconstructionArchitecture diagram, annotated modules, data flows, threat model, design decisions, synthetic fixtures, permitted fork or clean-room implementationMakes system mechanisms understandable without implying access to private code or confidential systems.
4. Three learning pathwaysBeginner worked example; intermediate modular adaptation; advanced cross-system evaluation; separate branches or clearly versioned foldersPermits accessible entry without sacrificing advanced research or architecture detail.
5. CI/CD and release assuranceTest plan, peer review, dependency and secret checks, SBOM where appropriate, build provenance, deployment hash, approvals, rollbackLinks claimed controls and intended features to verified artefacts and releases.
6. Operational outcomesAccessibility, accuracy, security, reliability, privacy, fairness, incident and performance indicators with dated baselinesDistinguishes implementation from evidence of success, failure or inconclusive findings.
7. Publication and continuityPaper type, research question, methodology, results, counter-evidence, limitations, data/code availability, contributor attribution, revision and correction logSupports reproducibility, professional review and long-term evidence preservation.

Difficulty pathways and repository architecture

Both arrangements are supported: one project repository with beginner/, intermediate/ and advanced/ directories, or three separately versioned code repositories where licensing, deployment, ownership or independent releases justify separation. In either approach, all levels should link back to the same source paper, upstream version and project identifier.

PathwayExample scopeEvidence/acceptance gate
BeginnerRead the architecture, trace a request, run a minimal local example with synthetic data, annotate a small component, explain trust boundaries and licensing.Reproducible setup, understandable diagram, basic validation tests, accessible README and no exposed secrets.
IntermediateModular redesign, automated tests, least-privilege configuration, dependency inventory, CI, error handling, accessibility and privacy review.Repeatable test suite, documented change comparison, dependency review, explicit acceptance criteria and release notes.
AdvancedMultiple services or agents, provenance-aware evaluation, independent assurance, resilience and counterfactual comparisons, authorised deployment telemetry.Threat-model and risk review, independent replication, baseline versus deployment results, governance evidence and incident/rollback plan.

Suggested skeleton: README.md, LICENSE, CITATION.cff, sources/upstream.md, papers/, docs/architecture.md, docs/compliance-matrix.md, docs/threat-model.md, beginner/, intermediate/, advanced/, tests/, evidence/, .github/workflows/, CHANGELOG.md and SECURITY.md. Actual files are created when an individual project starts; this is a reference structure, not a claim they already exist.

Existing Manifold Grace repository · GitHub repository duplication guidance · Template repository guidance.

Project families for legal, security and public-interest evaluation

Use the subject filter to explore candidate projects. Each family can support three difficulty levels and multiple distinct paper forms; no example is presented as a completed build. Additional interdisciplinary candidate families (16–23) appear below.

01 · Policy-to-code compliance mapper

Trace requirements to specific controls, tests, evidence owners and exceptions, with jurisdiction and effective-date versioning.

Compliance matrix · control traceability · legal update history

02 · Software supply-chain assurance

Trace source releases, components, dependency manifests, attestations and intended versus deployed artefacts.

SBOM · build provenance · CI/CD checks

03 · Secure web service reference

Study authenticated application architecture, input validation, access control and safe logging using non-sensitive synthetic examples.

Threat modelling · OWASP ASVS · security tests

04 · Privacy and consent lifecycle

Model personal-data collection, minimisation, retention, purpose limits, subject access, deletion requests and audit trails.

Data map · lawful basis review · rights workflow

05 · AI evaluation and safeguards harness

Compare model versions against labelled, consent-aware test sets; log uncertainty, failures, human decisions and corrections.

Model cards · evaluation reproducibility · oversight

06 · Agent governance and authority boundaries

Study tools, delegated permissions, approvals, policy checks, provenance and safe exception handling across agents.

Least privilege · tool audit · human escalation

07 · Historical publication and claims index

Preserve original source dates, revisions, stable links and evidence comparisons across publishers and decades.

Temporal provenance · omissions · corrections

08 · Accessible public-service workflow

Test how a digital service handles alternative formats, reasonable adjustments, human review and administrative corrections.

WCAG · equality considerations · process conformance

09 · Defensive incident response and recovery

Evaluate detection records, escalation, containment governance, backup verification, incident notices and recovery outcomes in a controlled simulation.

Resilience · continuity · post-incident review

10 · Procurement and vendor risk register

Compare contract requirements, suppliers, licensing, geographic data custody, security evidence and exit strategies.

Third-party assurance · contractual risk · interoperability

11 · Data lineage and knowledge integrity

Trace dataset origin, transformations, permissions and downstream model or decision outputs with independently testable evidence links.

Dataset cards · provenance graph · contradiction detection

12 · Evidence-led organisational observability

Compare documented authority, procedural decisions, independent oversight, correction handling and public-interest reporting.

Institutional process · due process · neutral audit

13 · Configuration and deployment drift monitor

Compare approved build/configuration, pre-release tests and observed production behaviour using authorised read-only monitors.

Version comparison · drift alerts · outcome metrics

14 · Safeguarding and accessibility assurance

Evaluate protective workflows, reporting routes, escalation safeguards, privacy controls and accessible support across service users.

Duty of care · confidential handling · human review

15 · Cross-institution interoperability sandbox

Test how corrected evidence transfers between systems without flattening uncertainty, losing original dates or exposing private records.

APIs · schema migration · integrity tests

Academic and professional paper options

Publication type should be selected to fit the question and evidence, not the other way around. Multiple outputs may share a project, provided their contributions and source records are distinguishable. These are potential formats, not claims that publication has occurred.

Paper or output typeMost useful whenRequired distinctions
Literature review / systematic or scoping reviewMapping established research, contradictions, gaps and breadthSearch dates, inclusion criteria and systematic status documented accurately.
Conceptual / theory paperDefining a framework, mechanism, taxonomy or falsifiable propositionSeparate conceptual contribution from empirical validation.
Architecture / design-science paperExplaining system components, design requirements, interfaces and trade-offsWorking artefact, proposed architecture and tested system clearly distinguished.
Methods / measurement paperIntroducing an evaluation protocol, rubric, dataset or measurement instrumentValidity, reliability, error analysis and replicability addressed.
Empirical experimental paperTesting causal or performance hypotheses with controlled comparisonsPredefined variables, baselines, limitations and negative findings.
Replication / reproducibility studyTesting whether a published result or code artefact can be reproducedOriginal experimental conditions and deviations reported.
Comparative benchmark paperContrasting models, architectures, controls or versionsFair datasets, environments, baselines and uncertainty.
Software / research artefact paperPresenting a reusable, documented implementationRelease, licence, dependencies, tests and citation details.
Security / assurance evaluationEvaluating authorised defensive controls and risk reductionDefined authorisation, safe scope, evidence and responsible reporting.
Legal doctrinal / regulatory analysisInterpreting duties and comparing rules across jurisdictionsOriginal instrument dates, applicable law and expert validation.
Policy / governance analysisComparing alternative oversight, accountability or compliance designsExplicit normative assumptions and implementation trade-offs.
Case study / longitudinal studyUnderstanding a bounded system or changes over timeSource provenance, case selection, chronology and competing explanations.
Ethnographic / qualitative studyUnderstanding user, worker or institutional experiencesConsent, safeguarding, sampling, positionality and confidentiality.
Participatory / action researchCo-developing and assessing a real interventionParticipation, feedback, governance and staged evaluation.
Position / perspective / commentaryArguing for a research direction or questioning accepted assumptionsArgument and proposed validation distinguished from established results.
Registered report / preregistered protocolPublishing proposed questions, design and analysis before resultsPreregistration does not imply completed experiments or journal acceptance.
Technical report / audit and assurance reportDocumenting professional testing, findings and corrective recommendationsScope, criteria, access, evidence status and reviewer limitations.
Education / curriculum or practitioner guideTeaching architecture with segmented worked examplesOriginal-source attribution, learning objectives and safe exercises.

Interdisciplinary social research, humanities, creative arts and public responsibility

Research extension recorded 13 November 2026. Study code, models and deployed artistic systems together with their social context: what is represented, how audiences interpret it, who benefits, whose labour and cultural knowledge are recognised, and whether platform design influences autonomy, dignity, participation or access. The following are candidate project families and evaluation designs, not completed experiments, proven claims about organisations or an assertion that one universal definition of artistic quality exists.

Interdisciplinary project familyHumanities / social-science connectionsArchitecture and reproducible evaluation
16 · Artistic-quality blind spots in multimodal modelsAesthetics, musicology, art history, philosophy, cultural studies, psychology of perceptionCompare image, audio and text models using consented or appropriately licensed culturally diverse works. Record model versions, task prompts and scoring rubrics; compare with qualified, plural human assessments, uncertainty and disagreements. Analyse technical fidelity, expressiveness, contextual appropriateness and originality separately.
17 · Representation, dignity and dehumanisation in AI-mediated artSociology, ethics, visual culture, critical media studies, anthropology, disability studiesAudit stereotype reproduction, erasure of human agency, decontextualisation and disrespectful portrayals through defined qualitative coding, diverse review panels, alternative interpretations and accessible participant protections.
18 · Persuasion and manipulation through creative media systemsCommunication research, behavioural sciences, rhetoric, social psychology, democratic participationStudy disclosure of synthetic media, persuasive framing, recommendation and engagement ranking with ethically approved benign materials. Measure recognition, informed choice and user autonomy without optimising manipulative techniques.
19 · Creative labour, attribution and cultural economicsLabour sociology, intellectual property, economics, cultural policy, media industriesTrace permissions, provenance, acknowledgements, compensation structures and creator controls through model training, publication, recommendation and licensing workflows; compare documented policies with observable outcomes.
20 · ESG and organisational responsibility in creative AIBusiness ethics, sustainability studies, governance, corporate accountability, cultural economicsEvaluate environmental claims, infrastructure/resource reporting, social accessibility, creative-worker impacts and governance evidence separately; compare public claims with dated, independently verifiable disclosures. Do not treat ESG labels as proof of performance.
21 · Cultural heritage and historical-aesthetic fidelityHistory, archaeology, architecture, linguistics, conservation and museum studiesCompare AI descriptions or recreations with dated sources and qualified curatorial assessments; preserve uncertain provenance, distinguish artistic interpretation from historical reconstruction, and evaluate whether minority traditions are omitted.
22 · Safeguarding, consent and synthetic identityChild protection, social work, law, media ethics, human-computer interactionReview consent, impersonation safeguards, likeness and voice rights, age-appropriate disclosure, appeal procedures, data minimisation and independent incident correction using controlled synthetic exemplars only.
23 · Creative discovery and visibility across communitiesMusic, literature, linguistic diversity, cultural sociology, information scienceTest recommendation-system exposure across languages, genres, regional artistic traditions and independent creators using documented datasets, eligibility criteria, confounders and fair comparisons.

Distinguishing artistic quality from measurable harm

Artistic value is multidimensional and can differ by period, culture, genre, tradition and audience. Aesthetic disagreement is not automatically algorithmic bias, and statistical popularity is not itself evidence of quality. Likewise, a polished generated output can still contain cultural errors, erased context, misleading provenance or unauthorised appropriation. Research should differentiate:

DimensionExample evaluation questionMethodological caution
Technical executionIs the output coherent, legible or technically well formed for its intended medium?Technical polish does not establish artistic or cultural merit.
Expression and intentionalityDoes the work convey a discernible expressive aim within its context?Do not infer the model possesses a human artist's intentions.
Creativity and originalityIs the result novel in a defined comparison set and meaningful for the task?Novelty, copyright status and aesthetic worth are separate questions.
Context and fidelityDoes it respect genre, historical tradition, language and source attribution?Distinguish intentional reinterpretation from factual error.
Audience interpretationHow do different communities and experts interpret the work?Report inter-rater disagreement and sampling limits; avoid a single universal score.
Human dignity and consentAre subjects represented appropriately, with lawful use and effective remedies?Assess harms using evidence and established duties, not aesthetic dislike alone.
Creative livelihoods and governanceAre contributors credited, protected and able to contest decisions?Separate company statements from independently observed practices.

Beginner, intermediate and advanced educational pathways

Beginner — explain a creative AI system and its social assumptions

Read an original model/paper with its publication date; inspect an openly licensed example with the exact upstream commit; diagram input, transformation and output; compare its descriptions of a small consented/public-domain artistic corpus with contextual notes; write a reflective ethics and aesthetics commentary. Output: annotated architecture, source register, qualitative coding guide and teaching notebook.

Intermediate — reproduce a cultural and quality evaluation

Build an authorised reproducible scoring harness, stratified examples across cultural traditions and a human-review protocol. Test whether ranking or captioning changes when context, language or medium changes; retain disagreements, uncertainty and negative results. Output: benchmark dataset documentation, comparison report, versioned code and optional methods or mixed-methods paper.

Advanced — cross-system accountability and longitudinal outcomes

Link lawful model access, deployment hashes, content provenance, creator permissions, exposure statistics, ESG disclosures and documented correction routes. Study longitudinal effects with suitable comparison designs and independent oversight. Output: architecture/design-science paper, social-impact or policy analysis, reproducibility package and qualified assurance report where evidence supports one.

Paper types especially relevant to the humanities and social sciences

Alongside the 18 paper forms already listed above, the most useful choices may include: critical interpretive essay, art-historical or comparative aesthetics study, media/discourse analysis, qualitative interview or ethnographic study, mixed-methods social impact study, participatory design and creative-practice research, historical reception study, creator-rights and cultural-policy analysis, and ESG materiality or impact assessment. Specify method, evidence type, review standards and ethical approval expectations for each; creative-practice research requires documentation of the actual practice, not simply generated outputs.

Safeguards and falsification

Register hypotheses in advance where suitable; include non-AI baselines, counterexamples and differences across genres and communities. Use independent, culturally competent reviewers; report uncertainty and disagreement; protect participant dignity and consent; avoid deceptive experiments without justified oversight. Do not assume that any observed stereotype or quality divergence proves intent or a particular organisation's misconduct. Retain author credit, source rights and a correction trail. ESG assessments should cover environmental, social and governance dimensions with separately evidenced metrics.

Authoritative starting references with original dates: UNESCO Recommendation on the Ethics of AI — adopted 23 November 2021; UNESCO Independent Expert Group on AI and Culture — 2025 report; NIST AI Risk Management Framework — released 26 January 2023. These support the research context; they do not validate the particular proposed experimental results.

Linked collapsible research questions · Existing research · Related news provenance index.

Project dossier and linked research records

Reusable project dossier — minimum fields

Project ID; concept and context; audience and intended use; project owner and responsible reviewers; current maturity (concept / design / prototype / tested / deployed); architecture and data-flow diagrams; upstream paper and repository/release/commit; licences and attribution; downstream derivation and code-diff record; beginner, intermediate and advanced learning objectives; relevant legal jurisdictions; compliance requirements; threat model and safeguards; privacy and ethical approval; test baselines; CI/CD evidence; expected and observed outcomes; limitations, counter-evidence and incidents; accessibility; stakeholder feedback; maintainers; dependencies and funding; reproducibility; release and review dates; correction and deprecation policy.

Five linked research records

Conceptual idea: problem, prior work, context and proposed contribution. Research project: objectives, ownership, work packages, milestones and resources. Paper initiative: prospective paper type, working outline, evidence needs and publication status. Contextual research question: scope, variables, counter-hypotheses and falsification plan. Productivity log: dated action, verified evidence link, why it matters and next step. Link all five by stable project ID. Create them as actual records when a specific project begins; do not label a proposal as a completed paper.

Review gates and independent oversight

Intake and relevance → upstream licence and citation verification → architecture review → privacy, safeguarding and regulatory assessment → threat model → implementation and peer review → synthetic-data testing → CI/CD checks → authorised deployment → post-deployment evaluation → independent replication → release, incident correction and retirement. Different risks warrant different levels of oversight; legal interpretation should be reviewed by a qualified specialist where needed.

Acceptance criteria and comparison

Every implemented level should have a measured baseline, testable functional requirements, security and accessibility checks, evidence of approved deployment, date-stamped results, known failure cases and a reviewer-visible change log. A passing test is not proof of legal compliance or absence of vulnerabilities. Compare like-for-like versions and retain negative results.

Licensing, intellectual property and lawful evaluation

Prefer documented open-source licences or the original author's explicit permission. Preserve attribution, notices, licence compatibility and dependency obligations; do not assume public repositories are licensed for cloning or modification. Never bypass authentication or paywalls, retrieve private code without authority, collect sensitive data without suitable safeguards, or test third-party systems without permission.

Standards and academic foundations

Source publication or release dates are preserved here rather than substituted with the date of this framework. Standards must be checked for revisions and jurisdictional applicability before using them as compliance criteria.

Original date / versionAuthoritative referenceUse in project evaluation
3 February 2022 (final v1.1)NIST SP 800-218 Secure Software Development FrameworkSecure development and supply-chain practices.
26 July 2024 (final)NIST SP 800-218A — generative AI community profileModel-specific secure development practices.
17 December 2025 (initial public draft)NIST SP 800-218 Rev. 1 draft (v1.2)Possible future revisions, explicitly not presented as final.
2021 (FAIR4RS principles; publication record)FAIR Principles for Research SoftwareFindability, accessibility, interoperability and reuse.
Version 5.0 (see release history)OWASP Application Security Verification StandardWeb application security requirements and tests.
2024 conference guidanceACM PADS artefact evaluationReproducibility, transparent experimental artefacts and review.
Current documentation (date varies)GitHub artifact attestation documentationLinking release artefacts with verifiable provenance.
Current documentation (date varies)W3C accessibility standardsAccessibility acceptance criteria and inclusive design.

Publication and verification status: this page is a design and research planning resource. Individual projects, code forks, compliance attestations, published papers and completed evaluations must be independently linked and verified before they are reported as delivered.