Manifold Grace
RESEARCH / OPEN QUESTIONS / EVIDENCE

Research Questions, Confirmed Gaps & Solutions

An evidence-first continuation record: what is known, who contributed, what remains unresolved, how a response could be tested and what future researchers can inherit.

Emerging Horizon · Discovery Directory · Dated source register

Before calling something a research gap

State a bounded claim, examine existing literature and implementations, reproduce evidence where possible, record competing explanations and credit earlier progress. A gap is confirmed only within the supported scope. If a solution is only designed or partially tested, publish the response with its real validation status rather than claiming the problem has been solved.

Types of questions include: descriptive, causal, comparative, methodological, architectural, evaluative, safety/regulatory, historical and counterfactual. Expert replies must be attributed to real named sources or explicitly marked as not yet obtained.

Question and solution register

Dated 5 November 2026 · Five curated examples · Original contributions and unresolved problems kept distinct

5 Nov 2026 · Confirmed local documentation gap — addressed

Research infrastructure and provenance

Research question
Can another investigator find the original research question, prior contributions, response, evidence and handover in one public place?
Evidence and scope
The earlier research-arenas and discovery pages offered categories and links but no single evidence-linked question-to-solution register.
Competing explanations
The earlier records contained research and design decisions, but readers still needed to join them together themselves.
Response / proposed solution
This dated register provides the missing navigable index and connects earlier records without replacing them.
Validation or limitation
Page and records can be inspected. This closes only the narrow publication-format gap; usefulness has not been independently assessed.
Continuation
Test with independent research readers; add versioned corrections and archive references.
Prior work, contributors and primary sources

1 Nov 2026 · Contributor-reproduced issue — upstream fix not verified

Benchmark answer normalization

Research question
Can a scoring function incorrectly process reference answers and credit a wrong generated response?
Evidence and scope
Issue #4502 contains a reproduction using HRM8K reference and predicted answers, including scoring reversals.
Competing explanations
Reference extraction, numeric normalization and extraction failure are competing error sources. The issue specifically distinguishes them.
Response / proposed solution
The contributor proposes separating reference normalization from predicted-answer extraction, failing closed on missing extraction and running regression cases.
Validation or limitation
Reproduced and reported by the named contributor. No upstream fix or independent retest is claimed by this site.
Continuation
Check issue and pull-request history, test against matched versions, document resulting score changes and credit all prior contributions.
Prior work, contributors and primary sources

22 Jan 2026 · Documented limitation — follow-on experiment proposed

AI agent time horizons

Research question
When do task-horizon estimates transfer to unfamiliar, context-rich, collaborative or high-reliability work?
Evidence and scope
METR researcher Thomas Kwa explains uncertainty, domain sensitivity and the difference between task difficulty and autonomous operating time.
Competing explanations
Task sampling, baseline conventions and the required reliability threshold may change how estimates should be interpreted.
Response / proposed solution
Proposed extension: compare domains and real-context conditions separately; measure human interventions and uncertainty, not just a single time-horizon estimate.
Validation or limitation
No new comparison has been performed here. Prior researchers' limitations and conclusions are preserved.
Continuation
Preregister population, task diversity and controls; conduct independent comparisons.
Prior work, contributors and primary sources

5 Nov 2026 · Candidate research question — novelty unconfirmed

Historical source revisions and claim evidence

Research question
Can source snapshots, observation dates, corrections and source-version identifiers improve handling of changing evidence?
Evidence and scope
AVeriTeC already provides evidence-grounded question–answer pairs, source metadata and verdict categories.
Competing explanations
Existing claim-verification systems may already address parts of the issue. A prior-art review is required before claiming an unsolved field-wide gap.
Response / proposed solution
Proposed extension: explicitly link observation time, publication time, source revisions, correction history and adjudication status.
Validation or limitation
Proposal only. It is not a demonstrated new solution or independently established research gap.
Continuation
Map prior art, assess dataset permissions, preregister evaluation and compare with established approaches.
Prior work, contributors and primary sources

5 Nov 2026 · Confirmed local validation gap — still open

Independent review of the local claim pipeline

Research question
Does the prototype's claim comparison agree with independently obtained human reference labels?
Evidence and scope
The v1.1 audit records 12 fictional cases with zero completed independent reviewer annotations.
Competing explanations
Passing software tests supports reproducibility, not externally assessed semantic correctness.
Response / proposed solution
A blinded dual-review workflow and gating process exist. The remaining substantive solution requires actual independent reviews and documented adjudication.
Validation or limitation
The validation gap remains open; no external agreement or semantic-accuracy figure is asserted.
Continuation
Secure two independent reviewers; preserve original sheets, adjudicate, freeze labels and only then score the pipeline.
Prior work, contributors and primary sources

Reproducible engineering demonstration

A local eight-case synthetic test compared a value-only baseline with an explicit subject, scope and observation-time guard. Against authored expected labels, the baseline matched 2/8 and the guarded comparator 8/8. The test ran on 5 November 2026; this demonstrates a specific behaviour in a designed fixture, not independent model accuracy or a new field-wide discovery.

Reproduce the Python test · View the result record

Limitation: Typed synthetic claims only; different observation times require cautious classification, not a claim that either source is true. No natural-language extraction, reviewer agreement or deployment assessment was conducted.

Architectural alternatives and trade-offs

ArchitectureAdvantagesTrade-offs and safety considerations
Rule-based / typed comparisonPredictable and inspectable, inexpensive, clear abstention rulesLimited language coverage; safest in narrow, explicitly typed conditions
Hybrid extraction + typed reasoningBroader input handling with stable comparison layerParsing errors may propagate; require field-level tests and abstention
Evidence-first / provenance-awareTraceable citations, versioned claims and reviewer accountabilityHigher data and operational complexity; evidence quality still requires judgment
Agentic multi-step evaluationFlexible information gathering and tool orchestrationMore attack surfaces, nondeterminism and costs; requires access controls and human oversight

There is no universally safest architecture. Choose according to context, authority, expected harm, validation evidence and monitoring. Restrict autonomy when consequences or uncertainty are high.

Lifecycle risks and mitigation record

PhaseRiskControls and reference
Discovery and scopingFalse novelty, omitted prior researchPrior-art search, sources, contradiction review and credit
Data and annotationPrivate data exposure, label biasSynthetic or authorised data, minimisation and independent review
ImplementationFragile parsing and incorrect reasoningTyped interfaces, unit tests, edge cases and conservative abstention
BenchmarkingContaminated sets, broken graders, exaggerated capabilityFrozen baselines, versioned graders, withheld tests and reviewer adjudication
Deployment and regulationExcessive autonomy, unfair outcomes, insufficient accountabilityRole-based access, impact assessment, human approval, applicable legal analysis, monitoring and incident response
Publication and archivingMisattribution, irreversible overclaim and silent correctionsPersistent links, dated errata, integrity records, licensing and source credit

Reference frameworks: NIST AI RMF, UK AI assurance guidance, ISO/IEC 42001, and MIT AI Risk Repository. These guide evaluation and governance; this project makes no conformity or certification claim.

Archive and research handover

For each completed project, retain: research question, prior art, evidence, baseline and method, original contribution, unsuccessful attempts, corrections, alternative design rationales, risks, independent review status, version history, and an explicit next study. Preserve all contributors' work and attribution, including partial advances.

New conceptual research proposal: Equitable Positive-Sum Growth: Green, Amber and Red Research Zones specifies a testable method to compare win–win feasibility, uncertainty, distributional harms and alternative architectures. This is a proposed evaluation framework, not a claimed empirical economic finding.

New interdisciplinary research field proposal: Semantic Integrity & Public Communication Assurance presents eight documented-source studies, twelve labelled illustrative cases, twenty-eight discipline intersections and a testable human/AI publication assurance methodology.

Earlier decision trace · Earlier experimental roadmap · Compact dated source record

This is a curated living register, not a claim that every field-wide gap has been located. New evidence can close or reclassify an entry; publish the correction and its date.

Research questions: formation, information environments and conflict prevention

RQ · Does long-term moral formation influence peaceful conduct?

Test whether longitudinal differences in voluntary education, mentoring and professional socialisation predict measured dispute-resolution outcomes after accounting for selection, context and socioeconomic factors. A null or adverse result must be retained.

RQ · Do information environments affect escalation risk?

Compare source diversity, hostile framing and recommendation-system exposure with measured empathy, dehumanisation, misinformation and peaceful-settlement preferences. Separate changes in attitudes from evidence of real-world conflict prevention.

RQ · Can independent QA strengthen credible early warnings?

Compare provenance-based review with established assessments using blinded outcomes, precision, recall, documented harms, independent replication and uncertainty. Do not label preliminary hypotheses as proven gaps.

Methods, contribution pathways and source framework.