Manifold Grace

RESEARCH & KNOWLEDGE · INTEGRATED SOCIETAL INTELLIGENCE

Public Authority, Integrity Gaps & AI-Assisted Reporting

Research framework for comparing the rules institutions publish with evidence of how those rules operate. The central proposed AI application is to help researchers and members of the public assemble, verify and report substantiated discrepancies—not to label people as corrupt.

Migration, fair treatment & cooperation paper initiative · Government directory · Public systems research · Dated activity record

Statistical basis and limits

85%Conflict-of-interest regulatory criteria fulfilled, average across OECD member and partner countries, 2025.
42%Practical criteria fulfilled, average across the same countries, 2025.
43 ppDifference between the two percentages; not a rate of corrupt officials or decisions.
£55–81bnUK public-sector fraud and error combined, estimated for 2023–24, with uncertainty.

Primary evidence: OECD Anti-Corruption and Integrity Outlook 2026, Chapter 4 and National Audit Office overview of fraud and error in 2023–24. The OECD gap is a multi-country average; it cannot be assumed to represent the UK specifically. The NAO estimate includes fraud and error, not merely deliberate wrongdoing.

Evidence standard: A role, social connection, donation, apparent generosity, disagreement with government or statistical anomaly is not evidence of misconduct. The AI must record uncertainty, seek alternative explanations, avoid profiling protected groups and require human verification before external reporting.

Roles and records to examine

Institutional settingDiscretionary rolesEvidence sourcesIndependent checks
Central governmentMinisters, accounting officers, senior officialsOfficial responsibilities, spending, declarations, published decisionsNational Audit Office; Parliament; departmental governance
Agencies and regulatorsDirectors, boards, case decision-makersAppointments, registers, enforcement records, annual reportsRelevant statutory appeals, oversight and audit functions
Local authoritiesCouncillors, officers, planning and procurement decision-makersInterest registers, planning decisions, contracts, auditsMonitoring officers, standards processes, external audit
Public servicesCommissioners, delegated managersCase criteria, complaints, service and purchasing outcomesSector regulators, inspectorates and ombudsmen
Private suppliersDirectors, beneficial owners, contract counterpartiesContract notices, ownership, related-party disclosuresProcurement authorities and appropriate regulators

Official directories: UK government organisations; ministers; Companies House; Contracts Finder. Any examination should use officeholders and rules applicable at the date of the decision.

Proposed AI evidence-to-reporting workflow

StageAI assistanceRequired safeguard
1. Capture obligationsExtract statutory duties, policies, eligibility rules and effective datesQuote and link to authoritative primary documents
2. Capture implementationOrganise dated public records, audit findings and comparable decisionsPreserve provenance, omissions and confidence limits
3. Gap detectionCompare obligations against documented procedures or outcomes; flag missing declarations and inconsistent decisionsA flag means review needed, never guilt
4. Statistical testingCompute rates, baselines, disparities, uncertainty intervals and sensitivity checksControl for case mix, selection effects and missing data
5. Contradiction testingIdentify documents that weaken or contradict an initial hypothesisRequire alternative explanations and falsification attempts
6. ReportingGenerate factual, source-linked summaries for the appropriate oversight channelHuman approval, privacy screening, neutral language and appeal/correction procedures
7. Follow-upTrack responses, corrections and outcomes over timeVersioned logs; no automated public accusations

The AI component is a proposed design, not a trained or deployed system. Its first evaluation should use synthetic or anonymised cases with known ground truth, measuring extraction precision, false-positive rates, evidence coverage, reviewer agreement, fairness and correction handling. Model training should not use unverified accusations as factual labels.

Relevant safeguards: UK Algorithmic Transparency Recording Standard; UK Data and AI Ethics Framework; NAO: Using data analytics to tackle fraud and error.

Falsifiable research questions

H1 — Are declarations and oversight requirements actually fulfilled?

Measure the proportion of required declarations submitted, verified and addressed against applicable requirements; test whether missingness arises from data availability or actual non-compliance.

H2 — Are comparable discretionary decisions treated consistently?

Compare similarly situated cases after adjusting for relevant lawful factors and documented reasons; examine counterexamples and decisions inconsistent with the hypothesis.

H3 — Do documented interests predict decision outcomes?

Use verified and legally relevant affiliations only; test with controls, robustness checks and independently reviewed case records. Association is not automatically causation.

H4 — Do subsequent donations correlate with influence or restitution?

Compare timing, recipient organisations, documented benefits and measurable remedies; avoid inferring intention from philanthropy alone.

H5 — Do private bodies exercise authority outside lawful mandates?

Evaluate specific rule-making and enforcement against legal delegation, consent and jurisdiction; distinguish lawful civil society activity from coercive unlawful conduct.

Research and implementation status

Evidence source review and research design documented: 11 November 2026. The figures above are aggregate context, not results from original case-level tests. No allegation register has been populated, no live reporting endpoint has been built and no machine-learning model has been trained. Further validation must be documented before interpreting any individual institutional case.