RESEARCH OPERATIONS · 5 NOVEMBER 2026 · WORKING ROADMAP
Active Research: Evidence-Ready Priorities and Next Deliverables
This page is a navigational and decision-support layer over existing Manifold Grace research. Its purpose is to reduce repeated setup and context switching, identify the next demonstrable research output, and support focused independent investigation. Original pages, submission dates, experiments and deferred investigations remain authoritative and untouched.
Scope and confidence: This is a qualitative, provisional prioritisation, not a measured stochastic forecast, task-completion claim or promise of automatic research. Value and evidence readiness are estimated editorial categories. Uncertainty and effort are relative planning judgements; no numerical probabilities, invented returns or completed experiments are asserted.
Uncertainty-aware prioritisation rather than a fabricated score
1. Evidence readiness
Can the next claim be verified with accessible records or a synthetic test?
2. Practical value
Does it create a reusable research, career, teaching or policy artefact?
3. Complexity and uncertainty
What expertise, compute, sources and independent checks are still missing?
4. Opportunity to reuse
Will the result resolve several open questions without overgeneralising?
Prioritise high-readiness, reusable outputs; maintain lower-readiness hypotheses without pretending to have completed their investigations. Update rankings when new evidence arrives. A future empirical prioritisation model could use observed task completion, effort and failure distributions, but this page does not yet have those data.
Next bounded action: Choose a single published policy proposal, identify jurisdiction, affected parties, powers, assumptions and a verifiable decision test.
Output: One bounded policy-evaluation case
Guardrail: Regulatory authority and policy effect require independent checks.
Rather than expanding every domain simultaneously, combine the first, second, third and tenth streams into one bounded, inspectable demonstration. The proposed bundle has four reusable parts:
Synthetic evidence pack: 15 fictional, non-sensitive claim/source pairs, with several conflicting dates, changes of definition and missing references. Make clear they are illustrative rather than real institutional findings.
Machine-readable claim ledger: statement, source, source version, timestamp, explicit versus inferred relationship, supported/disputed/unknown label, review status and correction history.
Reference rubric: evidence coverage, contradiction correctness, temporal reasoning, abstention when uncertain, authorisation, latency and computation estimates when measured.
Transparent test demonstration: compare a basic summariser with a deterministic evidence-aware checker; only later test actual model agents with comparable resources and independent adjudication.
These are specifications for a first bundle, not claims that the datasets, code, experiments or expert reviews have already been produced. The existing evaluation protocol supplies the methodological constraints and the research lab provides an appropriate educational context.
First reusable artefact: synthetic claim test pack
Version 0.1 contains 15 fictional claim pairs with reference relation labels, rationales and explicit synthetic provenance. The examples cover contradiction, temporal compatibility, changed rules, scope and metric mismatch, missing evidence, formal-proof overclaims, uncertainty, attribution and dependency review. These are illustrative reference annotations, not an empirically validated benchmark. Open the machine-readable JSON test pack.
Starter scoring: evaluate relation classification accuracy per category, false contradiction rate, abstention on ambiguous cases and explanation faithfulness; record human disagreements, model and prompt versions, tool permissions, latency and measured resource use. Do not attach measured accuracy or a confidence interval until actual repeated trials have been performed.
November 2026 frontier research — evidence-led extensions
This section updates the existing execution roadmap rather than adding a second research queue. It records selected 2026 technical developments, distinguishes verified primary-source evidence from unresolved claims, and assigns one bounded next output per research strand. The priority ordering is qualitative, based on evidence readiness, reuse and practical value; it is not a stochastic prediction or an ASI arrival forecast.
Primary-source results: ARC Prize reports sharply different ARC-AGI-3 scores for the same model under different memory/harness configurations, with measurable resource costs. Performance under a specific harness is not proof of general superintelligence.
Compare unassisted, memory-assisted and coordinator-assisted configurations on identical held-out tasks, with equal compute and equal wall-clock comparison tracks.
Versioned experiment specification and scorecard: model, harness, prompt, context retention, budget, accuracy, latency and cost.
Primary-source example: Google DeepMind's AlphaEvolve couples language-model proposals with automated evaluators and evolutionary search for algorithm optimisation. It does not establish unrestricted recursive self-improvement.
Build a bounded toy optimisation task with immutable reference tests, a proposer, evaluator and correction log. Measure gains against random search and no-feedback baselines.
Reproducible improvement-loop demonstration with failure and evaluation-cost records.
Google DeepMind's 2026 From AGI to ASI discusses scaling AGI, paradigm shifts, recursive improvement and multi-agent collectives as possible pathways; these are research hypotheses, not a confirmed chronology.
Compare each proposed pathway with its measurable prerequisites and distinguishing falsifiable evidence; mark unknowns.
Four-pathway evidence matrix, not a predicted date of ASI.
Mixed: autonomous science, federated clinical learning, physical robotics and repository migrations have varying levels of documented deployment; broad numerical claims in secondary summaries require study-specific verification.
Choose one publicly documented implementation with an available technical report and extract architecture, actual validation, limitations, safety controls and human involvement.
One bounded deployment case study; do not generalise claims across all hospitals or laboratories.
Claim verification and terminology safeguards
Supported: GPT-6 Astra exists; ARC Prize publishes ARC-AGI-3 results and distinguishes Standard from Provider Adapter harnesses; DeepMind published the From AGI to ASI report and AlphaEvolve. Qualified: “System 2” is an analogy for deliberate computation, not evidence that every model implements human dual-process cognition. Reinforcement learning, search, evaluation and externally structured agent loops are distinct mechanisms. Unresolved or over-broad: AGI as an agreed imminent milestone, an asserted universal ASI threshold of “ten thousand experts for ten years”, blanket 95% medical diagnostic accuracy, and claims that static evaluation has become obsolete. Static fixtures remain useful alongside dynamic environments, adversarial tests and longitudinal assessments.
Specific benchmark caution: ARC Prize's GPT-6 Astra comparison reports 62.7% under a Standard harness at maximum reasoning versus 99.9% with a Provider Adapter at high reasoning. The configuration changes the interpretation. Comparison should retain the task partition, memory policy, compute budget, source timestamp and external-tool permissions.
Primary-source reading links (as checked 5 November 2026)
Added 5 November 2026 as a source-qualified extension to the existing roadmap. Proposed outputs are not recorded as completed experiments; original contribution history and deliberately deferred investigations remain unchanged.
Prototype-to-evaluation execution sequence — November 2026
Decision: Use the existing Research Systems Workbench as the interface boilerplate, not as evidence that advanced intelligence has been achieved. Advance in bounded, verifiable increments. Each milestone links to an authoritative source page, a live development workspace or an existing synthetic fixture. This supplements the frontier research priorities above rather than replacing or duplicating them.
Evidence boundary: The Lovable project has been created, its agent reports the build finished, and source paths for the architecture canvas, simulation, evaluation, feature incubator and Python Lab are present. Those facts do not yet verify their interaction tests, deployment security, API connectivity or performance. The Python experiment kit was prepared separately; an attached local archive is not a deployed service. No paid hosting, third-party database or external API is connected here.
Rank / strand
Evidence readiness
Practical value
Next verifiable test
Publishable or reusable output
Existing reference
P1 · Working prototype acceptance
High: created Lovable project and page files
High: exposes integration defects before further development
Test all six routes on mobile and desktop; create and edit a graph, preserve node types, export JSON, run offline simulation, inspect traces and verify the approval gate never approves itself. Record failure cases.
Functional acceptance matrix, screen captures, bug list and version/commit identifier
High: 15 labelled synthetic claim pairs already recorded
Very high: reusable across multiple research and feature proposals
Run a deterministic checker and a basic model baseline on the same fixture. Evaluate relation labels, unsupported conclusions, evidence gaps, abstention, mistaken contradictions and trace completeness. Report denominators and limitations.
Versioned test report, result files, baseline confusion matrix and failed-case explanations
Medium: Python kit drafted separately; live integration unverified
Very high: portable backend and independently testable workflows
Specify OpenAPI contracts from typed claim, graph, trace and approval models; create a local FastAPI service with pytest and bounded asyncio jobs; verify that the TypeScript UI cannot silently execute arbitrary code or bypass approval.
Python repository scaffold, tested API contract, example JSON requests and test logs
High for documented proposals; medium for independent implementation
Very high: turns prior contributions into identifiable demonstrations
Select one original documented proposal (historical as-of retrieval, provenance, claim-to-evidence coverage or consent-gated evidence access). Record original submission separately; compare an independent prototype against alternatives and make a reproducible test.
Proposal-to-requirement traceability matrix, mock-up, functioning proof of concept and evidence-qualified attribution note
Medium: GitHub workflows available; prototype pipeline not confirmed
High: reduces repeat manual checks as research expands
Keep the prototype in a separate versioned repository, add lint/typecheck, pytest, unit/integration tests, dependency checks and protected preview builds. Fail the release gate on missing approvals or failed tests.
Green pipeline evidence, branch protections, staged deployment checklist and rollback trial
Medium: protocol established; cross-model experiments not run
High for scientific credibility after baseline works
Compare simple, tool-using, multi-agent and independently assured architectures at matched task difficulty, resources and approval constraints; include failure injections, concurrency and test-time compute.
Replicable comparative evaluation with observed costs, uncertainty and negative results
Specified: requirements, sources, assumptions and expected outputs explicitly versioned.
Executable: actual code and machine-readable inputs exist; demonstration is distinguished from simulation.
Tested: repeatable automated tests pass and observed failures are recorded.
Evaluated: independent comparisons, known limitations, costs and safety constraints documented.
Published: report links back to original dated proposals, methods and reproducibility artefacts; no unsupported adoption or AGI/ASI claims.
Updated 5 November 2026 following review of the new Lovable workbench and existing research registry. Rankings are qualitative, adjustable and not represented as statistically estimated outcomes. No previous research content, attribution record or deferred reading has been changed.
Progression and publication gates
Mapped
Scope, terminology, domain and links to existing research recorded.
Specified
Hypothesis, evidence needs, limitations and proposed measures documented.
Prototyped
Actual executable artefact exists and inputs/outputs are reproducible.
Evaluated
Independent checks, failures, costs and uncertainty recorded.
Published finding
Claims supported by the documented evaluation, not by the existence of a draft page.
Preserve source versions, original contributor records and as-of dates across all stages. A feature request remains a feature request even if a related capability appears elsewhere later; influence and independent invention require separately verified evidence.
Intentionally deferred research
The AI-generated mathematics audit, close reading of Wiener's Cybernetics, extensive thirty-domain validation and major external implementation work remain reserved for later specialist investigation, consistent with earlier decisions. Their reference pages stay accessible, but they are not marked completed, automatically advanced or submitted as new feature requests.
Milestone evidence and researcher handoff
Findings now recorded: synthetic verification progressed from an abstaining baseline (v0.3), through familiar-case heuristic improvement (v0.4), weak transfer under new wording (v0.5), typed claim contracts (v0.6), a restricted extractor (v0.7), and frozen-parser challenge testing (v0.8). The principal bottleneck isolated so far is extracting defensible claim structure from unfamiliar language, not simply adding agents or dashboards.
Immediate next test: freeze the current pipeline, obtain a genuinely independent double-annotated new case set, then compare extraction coverage and downstream reasoning separately. Treat the 24 authored v0.8 challenge pairs as exposed development evidence, not a fresh blind validation set.
v0.9 execution checkpoint (prepared, not validated): a standalone Python reviewer kit now supplies 12 fictional unlabelled pairs, independent Reviewer A/B CSV forms, an agreement calculator and an adjudication template. Local regression status: 101 passing tests; independent completed reviews: 0 of 12. The next external dependency is recruiting two reviewers and retaining their separate decisions before calculating any new model score. See the v0.9 milestone and expandable researcher handoff.
v1.0 execution checkpoint (consistency, not new source accuracy): five documented metamorphic checks across 20 already-exposed fictional pairs met all 100 local behavioural invariants; 104 automated tests passed. The v0.7 parser and earlier benchmark results remain unchanged. This removes one class of implementation uncertainty, not the external dependency on two independent reviewers. See the single lab milestone table and expandable v1.0 procedure. The next publishable scientific claim still requires blind labels and a frozen evaluation plan.
v1.1 execution checkpoint (auditability): a reproducible read-only checker confirmed five frozen file hashes, 100/100 existing metamorphic invariants and 109 passing tests. Independent reviewers: 0/12 completed; semantic accuracy and deployment safety are not established. The single milestone table and expandable v1.1 decision guide show how researchers can adapt extraction, provenance, oversight and compute to their intended outcome. Next external gate: obtain two genuinely independent blinded review sets, adjudicate disagreements, then score the frozen extractor on new labels.
Independent-review handoff available: twelve previously prepared fictional cases are now shared without scenario tags or answer labels, alongside reviewer instructions and separate A/B annotation forms. No third-party reviews have been received. See the existing milestone record for an expandable decision guide; do not treat this as a new accuracy experiment.
External review outreach, 5 November 2026: Eight tailored invitations to academic researchers and one applied-methodology enquiry to Full Fact AI were sent individually. These are confirmed outreach actions, not reviewer acceptances or endorsements. No independent annotations have been returned or adjudicated. Next gate: obtain two separate blinded reviewer forms, record disagreements, freeze labels, and only then compare system accuracy. See the single existing milestones and methods record.
Outreach delivery clarification (5 November 2026): The earlier eight academic messages and Full Fact enquiry were submitted and accepted by the sending email service, but one academic invitation subsequently received a receiving-server rejection. A corrected invitation was sent to the researcher's newer professional address, as listed on a recent public CV; delivery and acceptance have not been independently confirmed. This is a correspondence-integrity correction, not a new research result. Independent completed reviews remain at zero. The next evaluated phase requires separate reviewer annotations, disagreement records and frozen reference labels before any comparative accuracy claim.
Researcher handoff: distinguish evidence status from outreach or implementation status
Record communication as attempted, sender-accepted, bounced, independently acknowledged, reviewer-accepted and completed; do not collapse those states. Record engineering as versioned fixture, observed result, failure mode, component hash and repeated test. Record science only when the evaluation population, independent reference labels, uncertainty, source provenance, metric definitions and limitations have been documented. A reviewer invitation alone does not count as independent evaluation, and correct program output alone does not establish source truth.
Architecture choice by observed fault: a failed contact channel needs contact correction, not a new AI component; parsing failures suggest extractor redesign and new challenge tests; mismatched typed results suggest logic or ontology revision; unresolved or missing evidence suggests human review and abstention; absent blinded reference labels blocks model-comparison claims. Preserve prior results and append corrections with dates and the evidence supporting them.
Architecture decision trace available: the existing AI Research Lab now illustrates how observed experimental failures lead to bounded design decisions and falsifiable tests. A machine-readable five-decision handoff makes the lessons reusable without duplicating the earlier milestone table. This records methodological synthesis only; independent semantic validation remains outstanding.
Research-operations synthesis prepared 5 November 2026. Iterative planning, not a finding of prior technical success; adjust proposed rankings as verified information changes.
Bounded research transfer and 360° scope
The Research Transfer, Scope & Completion framework consolidates closure criteria, ten phase-specific safety and evidence risks, ten coverage perimeters, and cross-links to earlier work. It does not replace the single milestone record or advance validation status.