Manifold Grace

Manifold Grace · Research development · 4 November 2026

Verifiable Advanced Intelligence

Evidence, causality, coordination and assurance across heterogeneous AI systems. This proposed research programme investigates whether claim-level verification, probabilistic and causal reasoning, procedural safeguards and continuous evaluation improve the reliability and generalisation of advanced agents.

Status: Conceptual proposal and evaluation design, not an implemented or validated ASI system. Concurrent execution or expert-level performance in a narrow task does not alone demonstrate artificial superintelligence. Original proposal dates and sources should be preserved in the linked contribution register; this page documents a subsequent synthesis, not an assertion of priority over independently developed ideas.

Twelve concurrent verification functions

The original five functions—historical fact-checking, visual claim extraction, definitional comparison, forecast tracking and contradiction assessment—are retained and extended, not replaced. Individual functions may run concurrently, but some require dependency ordering and independent review.

01 · Historical verification

Chronology and accurate historical claims

02 · Multimodal claim extraction

Explicit and implied claims from images, charts and text

03 · Definitional alignment

Equivalent meanings, non-equivalent taxonomies

04 · Forecast tracking

Predictions, assumptions, updates and observed outcomes

05 · Contradiction assessment

Genuine disagreement versus differences in scope

06 · Causal inference

Interventions versus correlations

07 · Counterfactual analysis

Alternate conditions and plausible consequences

08 · Stochastic modelling

Calibrated uncertainty and outcomes

09 · Procedural compliance

Rules, rights, required steps and exceptions

10 · Multi-agent consistency

Shared state, cross-agent coherence and conflict resolution

11 · Outcome verification

Authorised actions and measured effects

12 · Independent assurance

Reproducibility, audit trails and independent testing

Seven infrastructure layers and alternative architectures

Evidence fabric
Source archives · provenance · temporal retrieval
→
Knowledge models
Ontologies · causal graphs · uncertainty
→
Orchestration
Parallel tasks · dependencies · checkpoints
→
Assurance
Independent tests · audit · human control
  1. Evidence: multimodal retrieval, rights-aware archives and source versioning.
  2. Representation: temporal graphs, ontologies, claim objects and evidential provenance.
  3. Reasoning: symbolic, neural, causal, probabilistic and constraint-solving components.
  4. Execution: distributed workflows, tool brokers, event queues and recovery.
  5. Operations: authorised APIs, business process automation, simulations and scoped action triggers.
  6. Evaluation: process tracing, calibrated uncertainty, falsification, adversarial tests and independent review.
  7. Governance: identity, permissions, human approvals, rollback and incident accountability.

Architecture selection should be domain-specific: central orchestration for tightly regulated tasks; decentralised agents for independent research work; neuro-symbolic hybrids for rule-bound reasoning; causal-probabilistic models for interventions; digital twins for physical systems; and independent proposer–verifier–auditor designs for assurance-intensive environments. These patterns can be combined.

Related AI branches

Relevant, overlapping fields include neuro-symbolic AI, causal machine learning, Bayesian reasoning, multi-agent systems, agentic AI, scientific machine learning, knowledge representation, formal methods, distributed computation, computational neuroscience, process mining and business process automation (BPA/RPA). Computational neuroscience and neuro-symbolic AI are distinct fields; business process automation is not itself general intelligence.

Thirty domain-specific architectural profiles

Each row proposes a different application, specialist AI methods, enabling infrastructure and an assurance question. These are research candidates, not validated systems or claims of operational readiness.

ArenaSpecialist AI branchesInfrastructureIndependent assessment
1. Scientific discoveryScientific ML · formal reasoning · Bayesian optimisationResearch repositories, laboratory interfaces, reproducible workflowsReproducibility, novelty and independent experiments
2. Drug developmentBiological foundation models · causal inferenceClinical evidence graph, secure trials data, simulationSafety, trial validity and experimental confirmation
3. GenomicsStatistical genetics · graph AIPrivacy-preserving genomic knowledge systemsPopulation transfer and clinical validity
4. Materials engineeringPhysics-informed ML · optimisationMaterials databases and physical simulatorsMeasured properties and replication
5. Climate scienceEarth-system models · probabilistic MLGeospatial data fabric and climate digital twinsCalibration, hindcasting and uncertainty
6. MathematicsNeuro-symbolic AI · theorem provingFormal proof assistants and verified solversMachine-checkable proofs
7. HealthcareClinical decision support · multimodal AIInteroperable clinical records and consent controlsClinical outcomes, equity and clinician oversight
8. Public healthEpidemiological modelling · agent simulationPopulation data, statistical models and event feedsForecast calibration and intervention impacts
9. NeuroscienceComputational neuroscience · dynamical systemsNeural observations and research data standardsBiological plausibility and independent replication
10. Assistive systemsAdaptive AI · human-computer interactionAccessible interfaces and individual controlsUsability, autonomy and accessibility outcomes
11. EducationKnowledge tracing · pedagogical agentsCurriculum graphs and transparent assessment recordsLearning gains, fairness and validity
12. Public policyCausal economics · simulationAdministrative data and legislation timelinesDistributional effects and robustness
13. Legal procedureRule engines · legal NLPVersioned law and procedural evidence registryJurisdiction, rights and appealability
14. Welfare servicesBPA · event-driven decision supportCase records, rules engines and human reviewEntitlement accuracy and fairness
15. Education governanceCompliance AI · document reasoningAssessment provenance and adjustment recordsConsistent procedure and accessible challenge
16. PeacebuildingGame theory · participatory modellingConsent-aware evidence and scenario systemsConflict sensitivity and human oversight
17. Electricity gridsControl systems · digital twinsSensor telemetry and resilient grid simulationStability and operational safety
18. TransportationGraph optimisation · multi-agent schedulingNetwork digital twins and disruption feedsReliability, safety and equity
19. ManufacturingRobotics · predictive maintenanceIndustrial telemetry and process controlsFaults, QA and safe recovery
20. Water managementSpatiotemporal forecasting · anomaly detectionHydrological sensors and regulatory thresholdsContamination, calibration and alerts
21. Urban planningGIS · BIM · optimisationSpatial models, codes and scenario enginesAccessibility, compliance and lifecycle costs
22. Financial supervisionNetwork analysis · anomaly detectionAuditable transaction patterns and regulationsFalse positives and systemic risk
23. Labour marketsSkills ontologies · causal inferenceJob histories, occupational standards, skills graphsFair recruitment and mobility
24. Supply chainsNetwork optimisation · forecastingSupplier provenance and dependency graphDisruption resilience and continuity
25. Enterprise operationsBPA/RPA · process mining · agentsAuthorised APIs, workflow queues and rollbackCorrectness and exception handling
26. Privacy governancePolicy-as-code · data lineageIdentity, permissions and retention systemsLawful access, auditability and minimisation
27. Historical archivesTemporal graphs · entity resolutionVersioned archives and source registriesAuthenticity and chronology
28. Language and heritageSpeech AI · digital humanitiesCommunity-governed language corporaFidelity, consent and community authority
29. Journalism and public evidenceMultimodal verification · claim graphsSource archives and correction registriesCorroboration and provenance
30. Emergency coordinationGeospatial AI · stochastic operationsResilient incident data and human dispatch linksTimeliness, safety and accountable decisions

Research evaluation protocol

Compare four controlled conditions on the same new tasks: (A) foundation model alone; (B) model with retrieval and specialist tools; (C) coordinated agents without an independent evaluator; (D) agents plus the proposed claim-level verification architecture. Maintain comparable task selection, evidence access and compute budgets; also report latency and cost. Separate final-answer correctness from process quality and authorised action.

Evidence and semantics

Claim extraction precision/recall, citation coverage, source authenticity, temporal accuracy, contradiction classification.

Causal and probabilistic

Intervention identification, counterfactual robustness, calibration, uncertainty reporting and distribution shift.

Operational and social

Dependency consistency, task success, exception handling, compliance, harm prevention, accessibility and appealability.

Generalisation and efficiency

Held-out domains, expert baselines, human adjudication, reproducibility, time, cost and compute.

Use independent assessors and explicit failure analyses: agreement among agents sharing a model does not imply independent corroboration. Record original forecasts and later revisions separately. A diagram's layout may suggest chronology without explicitly claiming that older technologies disappeared.

Illustrative case: competing AI-development diagrams

Historical timelines, benchmark curves, ANI–AGI–ASI diagrams and forecast charts can be mapped into typed claim records: source, date, quote or spatial inference, definition, timescale, confidence and primary evidence. A system should distinguish disputed definitions, outdated projections, historical errors, compatible observations and genuine contradictions. Its output must keep unresolved claims visible rather than inventing consensus.

Four linked studies

  1. Cross-Representational Epistemic Verification: can systems reconcile claims in text, images and diagrams while preserving attribution?
  2. Multi-Agent Causal and Stochastic Reasoning: how do concurrently generated causal forecasts remain coherent?
  3. Domain-Specific Intelligence Architectures: which combinations of methods work reliably in which arena?
  4. Continuous Assurance for Advanced AI: does independent verification improve novel-task performance and reduce operational failures?

Pre-publication methodological development

This study treats verification as a testable capability-supporting mechanism, not as an established sufficient condition for AGI or ASI. The independent variables are orchestration design, availability of specialised tools, verification protocol and test-time computation; dependent outcomes include accurate claims, error recovery, procedural correctness, novel-task generalisation, safety and expense. Propose and assess narrower hypotheses before inferring general intelligence.

Operational hypotheses

  1. H1 · Evidence: claim-level provenance and cross-modal checking lower unsupported assertions relative to retrieval-only baselines on matched unseen source sets.
  2. H2 · Coordination: structured shared state and dependency-aware scheduling improve correct completion of tasks requiring several concurrent specialist analyses relative to uncoordinated agents.
  3. H3 · Causality: explicitly represented causal assumptions and intervention checks reduce mistaken causal claims relative to purely correlational synthesis.
  4. H4 · Assurance: a separately configured verifier and auditable correction loop improve reliability and reduce unsafe or unauthorised actions without prohibitive latency or computation.
  5. H5 · Transfer: improvements persist on held-out domains, unfamiliar diagram types, changed rules and newly sampled exceptions, rather than being confined to repeated benchmark tasks.

Experiment design: fair comparisons, including resource budgets

Randomise or balance task allocation; pre-register primary outcomes; hold underlying models, tool permissions and relevant source availability constant where the comparison requires it. Run repeated trials with recorded seeds and versions. Compare agent configurations at equal total compute budget as well as equal time budget, because parallel branches may consume more resources for the same task. Separately report performance-versus-compute curves. Use held-out test sets and blinded human expert adjudication for non-automatable claims.

Report per-domain results, not a single aggregate intelligence score. Include confidence intervals, disagreements between graders, reference-answer uncertainty, false positives and false negatives, and negative results. Stress-test shared-source errors, model-family correlated failures, unavailable evidence, conflicting policies, tool failures, distribution shift and adversarially misleading diagrams.

Typed claim and provenance record (illustrative schema)

{
  "claim_id": "example-001",
  "source_id": "archive-item-2026-001",
  "source_type": "diagram",
  "original_statement": "An explicitly extracted statement",
  "interpretation_type": "explicit | implied",
  "domain": "AI history",
  "published_at": "YYYY-MM-DD or unknown",
  "valid_from": "YYYY-MM-DD or unknown",
  "definition_id": "taxonomy-v1",
  "evidence_links": ["primary-source-id"],
  "relation_to_other_claim": "supports | contradicts | differs | unknown",
  "assessment": "supported | unsupported | disputed | uncertain",
  "reviewer": "human-or-independent-system",
  "revision_history": []
}

This record is a conceptual example, not an API standard or validated ontology. It distinguishes when an infographic explicitly states something from what a reader infers through arrows, position and visual sequence. Every correction should retain the original claim and its assessment history.

Cross-agent interaction protocol

Request with authority and scope
Planner builds dependency graph
Specialists check sources and models
Independent reviewer tests outputs
Approve, escalate or revise

Parallel branches should exchange provenance-bearing observations rather than merely free-text summaries. Escalate contradictions that remain materially unresolved. Any real-world operation requires scoped permissions, an identified accountable owner, explicit human approval where necessary, logging and reversibility where feasible. Tasks with safety-critical consequences need domain-specific professional oversight; no benchmark result by itself authorises deployment.

Worked teaching case: three conflicting AI timeline diagrams

Input: diagram A places an event in year X; diagram B places the same event in year Y; diagram C states that its dates are forecasts rather than observed events. Extraction: record each literal claim, diagram title, publisher, publication date and whether positioning implies rather than asserts a date. Reconciliation: check whether the labels refer to the same event; retrieve a dated primary announcement or historical source; classify A and B as compatible, incompatible or unresolved. Forecast: preserve C as a time-stamped projection and evaluate separately once an observable outcome exists. Output: a claim table with sources and uncertainty, not an averaged year or invented agreement.

Publication and attribution safeguards

Retain earlier feature-request records verbatim with original submission dates and available recipient responses. Date this synthesis as a new research-development record. Distinguish the author's documented proposal from existing related research and avoid claims of technical implementation, adoption, independent invention or causal industry influence without evidence. Do not reproduce third-party infographics in public pages without checking the licence, permission or applicable exception; a short original diagram or appropriately attributed link may be preferable.

Criteria to advance beyond the concept stage

Progress to an empirical report only after a task dataset, reference answer protocol, domain expert reviewers, preregistered measurements, error taxonomy, reproducible code or system configuration, and a resource-accounted evaluation exist. Release a model or service only after separate privacy, legal, domain safety and accessibility reviews. A stronger performance result is not automatically evidence of ASI: the claim would require much wider and more demanding cross-domain demonstrations.

Selected external foundations

Agent architecture and operational deployment

For ten agent topologies, original conceptual diagrams of parallel verification, hybrid workflow control and cloud–edge systems, and detailed security and evaluation considerations, see Agentic Architecture: From Automated Workflows to Verifiable Parallel Intelligence. These engineering patterns do not independently establish AGI or ASI.

Research · AI Research Lab · Evidence Verification Across Sectors · Information Science · AI Product Feedback · Feature Requests & Contribution Register

Research synthesis dated 4 November 2026. Additive development; earlier feature requests and dated records remain authoritative for their original wording.