Manifold Grace
RESEARCH CLOSURE · EVIDENCE PERIMETERS · SCIENTIFIC HANDOVER · 5 NOVEMBER 2026

Research Transfer, Scope & Completion

What has genuinely been learned, what remains unverified, how architectural decisions were reached, and how future investigators can extend the research without replaying or overwriting earlier work.

Original milestones · Architecture decisions · Execution roadmap

Research completion does not require perpetual perfection

A bounded research contribution is complete when its specified questions, methods, observations, limitations, risks, decisions, evidence provenance and handover are adequately documented. It need not solve every adjoining scientific problem or wait for all future research to conclude. Uncompleted independent validation, comparative architecture tests and production release must remain explicitly pending; completing a documentation phase does not convert those claims into verified findings.

This page is a research-transfer index and perimeter declaration, not a new benchmark report or a replacement for the original AI Research Lab. The underlying experimental records, reviewer packet, original feature proposals and previously dated work remain authoritative.

Audit: what the present record does and does not establish

Documented local engineering

v0.1–v1.1 offline, synthetic work; reported 109 passing tests and five preserved baseline component hashes in the v1.1 audit.

Documented limitation

v0.8: 12/18 end-to-end authored-label relation matches versus 18/18 supplied typed-input matches. Extraction is a candidate bottleneck.

Review material located

Twelve assistant-authored fictional cases and A/B blank forms are in the repository, with reviewer instructions.

Independent validation blocked

Zero of twelve candidate cases independently annotated at the documented checkpoint; no external agreement or novel semantic-accuracy result established.

Phase 6 is proposed

Baseline, hybrid extraction and evidence-first alternatives are candidates, not completed and compared implementations.

Operational safety unproven

No certified, independently assessed or production-ready general AI assurance service is claimed.

Correction to earlier handoff visibility: The reviewer package, including the twelve blinded fictional cases, is currently in the public repository. This establishes packet accessibility, not completion of review. As the test set is exposed and assistant-authored, independent reviewer annotation alone would still not constitute a fresh independent test population.

The original single milestone table details v0.1–v1.1 conditions and scores, while the five architecture decision records link each failure to its response and falsifiable release gate. These are intentionally not copied into a second results table.

Milestone learning: what makes the next architecture justifiable?

Observed failure or constraintDesign implicationWhat would change the conclusion?
Strong agreement on familiar lexical cases followed by weak transfer on unfamiliar wordingDo not optimise only for developer-authored examples; preserve test splits and introduce a structured extraction layer.Frozen-baseline comparison on a genuinely novel corpus with predefined metrics.
Typed-input reasoning performs better than end-to-end natural-language interpretationSeparate extraction from comparison, expose abstentions and inspect field-level error.Independent labels verifying that revised extraction improves novel inputs without worsening false certainty.
Software invariants pass but semantic validation is unavailableKeep code consistency, source truth, reviewer agreement and operational security as separate approval gates.Double-reviewed labels, adjudication, new test population, risk-weighted errors, and deployment assurance.
Multiple architectures are plausible but have not been comparedRetain A (frozen baseline), B (hybrid extraction) and C (evidence-first) as candidates without claiming superiority.Common evaluation set, cost/latency/safety measurements, error traces, controlled changes and independently auditable results.
Need for successor researchPreserve questions, rejected choices, decisions, tests, files, versions, incidents and uncertainty, not just pass counts.A new investigator reproduces a result and understands why each alternative was chosen or rejected.

Next bounded stage: validation before architectural comparison

  1. Freeze and identify source materials: confirm case IDs, source-file versions, model outputs, environment dependencies and hashes; preserve the v0.8 and v1.1 records unchanged.
  2. Check review eligibility: use the existing v0.9 12-case packet as a methodology pilot, not a new independent blind generalisation benchmark, because its authorship and exposure are documented.
  3. Obtain genuinely independent labels: involve two separate reviewers without sharing predictions, with consent, relevant expertise, a clear rubric and protected submission routes; log disagreements and unanswered cases.
  4. Construct a further non-exposed test set: independently sourced and controlled for licences, privacy, language, domain, scope and temporal complexity; freeze before model scoring.
  5. Evaluate baseline A, hybrid B and evidence-first C: keep conditions, endpoints, datasets, acceptance thresholds and compute accounting comparable; label any unimplemented candidate as proposed.
  6. Publish a stage report: pinpoint observed findings, negative results, trade-offs, safety limitations, unresolved questions and evidence-grounded architectural decisions.
  7. Close or transfer: mark the bounded stage complete, conditionally complete or blocked; preserve future hypotheses as a successor agenda without pretending they were tested.

Handover package minimum: research question and novelty boundary; dataset and permissions; code and dependency versions; original source records; reference labels and their independence status; test definitions and outputs; error analyses; architectural decision records; safety/legal review; contact and correction history; and a one-page successor decision tree. If any component is unavailable, mark it explicitly rather than reconstructing original evidence and calling it authentic.

Risk and mitigation by implementation phase

Risk assessment should follow the lifecycle, not appear solely as a final compliance paragraph. Each stage exposes distinct epistemic, technical, organisational and legal failure modes. These are risk hypotheses and recommended controls, not evidence of an actual breach or confirmed harm.

1. Problem formulation & historical surveyRisk and safeguard
Potential failure
Novelty inflation; omitted prior work; incomplete evidence
Controls to evaluate
Prior-art register; scoped question; provenance and opposing interpretations
Evidence boundary
Unverified novelty and inaccessible sources must remain open
2. Discovery, sourcing & intakeRisk and safeguard
Potential failure
Outdated sources; fabricated documents; non-consensual disclosure; source-selection bias
Controls to evaluate
Origin, publication/validity dates, access rights, source status, redaction review
Evidence boundary
Unpublished or private evidence is not automatically usable
3. Dataset and label designRisk and safeguard
Potential failure
Synthetic ground truth mistaken for reality; author-label leakage; demographic/linguistic undercoverage
Controls to evaluate
Independent sampling and blind labels; inter-rater disagreement; subgroup plans
Evidence boundary
v0.9 candidate cases are assistant-authored; no independent labels
4. Extraction and semantic interpretationRisk and safeguard
Potential failure
Paraphrase loss; wrong entities; temporal collapse; unsupported inference
Controls to evaluate
Separate extraction metrics from relation logic; abstention and error codes
Evidence boundary
v0.8 challenge results implicate the extractor, not every downstream rule
5. Typed reasoning and comparisonRisk and safeguard
Potential failure
Correct computation over incorrect premises; inconsistent ontologies; logical overreach
Controls to evaluate
Schema tests, contradiction/temporal tests, reason traces, source requirements
Evidence boundary
18/18 structured-input matches do not establish document understanding
6. Architectural comparisonRisk and safeguard
Potential failure
Test leakage; unfair baselines; unmeasured cost; model substitution
Controls to evaluate
Freeze datasets/configurations; same review set; latency/cost; rollback and reproducibility
Evidence boundary
Three proposed approaches have not been demonstrated comparatively
7. Agent orchestration and integrationRisk and safeguard
Potential failure
Permission creep; cascading tool calls; opaque delegation; errors accumulating over time
Controls to evaluate
Minimal scoped permissions, traces, human approval gates, rollback and stop conditions
Evidence boundary
Agentic extension is a candidate design, not evidence of reliable autonomy
8. Compliance, privacy and securityRisk and safeguard
Potential failure
Sensitive-data leakage; unclear operators; unaudited deployments; cross-jurisdictional duties
Controls to evaluate
Inventory deployments; classify purpose and applicable law; access control; DPIA/safety assessment where required
Evidence boundary
Audit obligations vary by role, sector, jurisdiction and risk
9. Independent evaluation & assuranceRisk and safeguard
Potential failure
Overfitted metrics; misleading pass counts; reviewer conflicts; false assurance
Controls to evaluate
Two genuine blind reviews; disagreement adjudication; new test population; acceptance thresholds
Evidence boundary
At last checked, 0/12 independently annotated; external validity unresolved
10. Publication, adaptation & decommissioningRisk and safeguard
Potential failure
Overclaiming; losing negative findings; outdated guidance; abandoned deployments
Controls to evaluate
Versioned decisions, public limitations, correction policy, succession pack, retirement checks
Evidence boundary
Published methodology can be complete while validation remains pending

What a defensible “360° AI research” perimeter requires

There is no valid numerical percentage for “everything historically, currently and in the future” without a finite, independently justified population of research topics and a coverage method. AI research is open-ended; future discoveries cannot be exhaustively catalogued in advance. Instead, the following ten dimensions identify coverage obligations and known blind spots. The site currently maps portions of these dimensions; it is not a complete, independently validated world census of AI research.

1. Temporal

Historical origins; as-of-date evidence; 2026 current state; future hypotheses

Limit: Do not infer a complete future or comprehensive historical census.

2. Disciplinary

Computing, mathematics, philosophy, linguistics, information science, law, policy, economics, sociology, safety and education

Limit: Subject affinity is not assessed achievement or specialist authority.

3. Sectoral

Public services, health, justice, finance, education, industry, science, media, infrastructure and civil society

Limit: Sector-specific obligations and empirical conditions differ.

4. Institutional

Individual, team, supplier, organisation, regulator, state and international coordination

Limit: Accountability can cross institutional boundaries.

5. Research lifecycle

Problem discovery, literature, evidence collection, ontology, design, experiment, review, integration, monitoring, archival transfer

Limit: Methods may be iterative; progress is not necessarily linear.

6. AI capability

Learning, prediction, language, multimodal, robotics, agents, reasoning, retrieval and simulation

Limit: Tool use and benchmark success are task-specific.

7. Evaluation

Functional correctness, semantic validity, external review, reliability, fairness, robustness, calibration, security and cost

Limit: No one pass rate substitutes for all dimensions.

8. Epistemic state

Verified record, developer-authored result, independent finding, provisional inference, hypothesis, unknown and disputed

Limit: Maintain provenance, confidence and competing explanations.

9. Governance

Identification, ownership, legal applicability, transparency, privacy, safety, monitoring, remedy and retirement

Limit: Do not assume every AI deployment has identical statutory requirements.

10. Professional transfer

Study themes, prior work, QA, published research, feature requests, CPD, professional bodies, collaboration

Limit: Document relevance without conferring titles, membership grades or credits.

Formal coverage reporting should therefore use bounded statements such as “34 of 40 defined historical topics have at least one primary source dated by X” or “8 of 10 selected disciplines have an expert-reviewed synthesis,” rather than an invented percentage of universal AI knowledge.

Additional phases and non-linear connections

Research can involve pre-discovery framing and terminology negotiation, rights and consent review, horizon scanning, historical archaeology, community co-design, epistemic audit, requirements and threat modelling, experimental construction, governance approvals, implementation, comparison, red teaming, stakeholder interpretation, correction, policy transfer, procurement, public communication, retirement and future re-evaluation. These phases can recur or proceed in parallel. The relation between phases should be encoded as dependencies, feedback loops, contradictions and evidence updates—not a single irreversible sequence.

Contextual lineage: studies, prior work, research, proposals and professional practice

This is a relationship index, not evidence that all listed programme modules were assessed, all feature proposals implemented, or membership grades awarded. It helps researchers reconstruct the relevant context without duplicating the original records.

Closure test and handover readiness

DeliverableCurrent defensible stateClosure criterion
Historical local milestone recordDocumented in AI Research LabStable links and archived source/fixture provenance
Architecture decision rationaleFive machine-readable ADRs presentEach important design response tied to an observed limitation and testable prediction
Independent reviewer workflowPacket and blank forms available, zero completed in recorded snapshotTwo genuine separate assessments and adjudication history retained
External semantic generalisationUnestablishedFrozen genuinely new independent set and agreed evaluation metrics
Comparative architecture resultsProposed, not performedComparable baseline/hybrid/evidence-first outputs with cost and risk metrics
Safety and regulatory deployment sign-offNot establishedRole-, sector- and jurisdiction-specific control evidence and authorised assessment
Knowledge-transfer publicationThis bounded index and linked originalsSuccessor can locate evidence, limitations, alternatives, next tests and missing materials
Global 360° inventoryCurated, open-ended and incompleteOnly claim coverage against predefined, dated and reviewable scope; maintain unknowns

Responsible completion statement: The research-methodology and historical handover deliverable can be shared as a bounded contribution now. Independent semantic validation, candidate-architecture superiority, production readiness and exhaustive global coverage remain unestablished. Their absence does not erase the documented learning; it defines the limits within which that learning may responsibly be reused.

Publication status: research synthesis and evidence navigation, 5 November 2026. No additional benchmark run, independent review, new scientific result or external endorsement is claimed by creating this page. Original dated artefacts are preserved.