From worthwhile questions to research that can continue.
AutoResearch brings question discovery, literature preparation, idea exploration, critical review, structural recombination and evidence communication into one research programme. Each module retains its own responsibilities and maturity.
Questions, sources and actions belong to one line of inquiry.
Connected research workstreams
Each workstream has its own responsibility, implementation status and entry. Their relationships are introduced here; each tool keeps its own version record.
The contribution-review and writing module in AutoResearch, connecting lens routing, independent critique, evidence profiles, and deterministic checks with research taste and author-guided writing.
A structured route through source discovery, record processing, staged screening, thematic synthesis, and citation review, supporting evidence preparation within AutoResearch.
InkTide brings inspected paper lessons, author evidence, original prose and figure design into a connected writing workflow, with a separate review of the result.
A map of responsibilities. The catalogue includes design protocols, local tools and external utilities; it does not imply a single deployed runtime.
Connecting the stages
Dashed connections show conceptual handoffs. The modules are not one deployed runtime.Swipe to inspect the full diagram.
Wenjie certificates provide value goals and discriminating envelopes; the Innovation Generator explores concrete mechanisms and subsequent actions.
Fusion can recombine parents, but its necessity predictions require experiments and independent review.
Research Hub records sources, frozen plans, observations and failures; LSMFC compiles manuscripts and figures from checked claims and facts.
Teaching and human understanding are parallel goals. The architecture keeps responsibilities independent; teaching satisfaction is not agent experimental gain, and the full teaching service remains unintegrated.
Literature Review prepares traceable sources; Themis Court examines contribution identity, evidence and objections. These are proposed handoffs, with runtime integration still to be verified.
Question discovery and research portfolios
Public question-design responsibilities; implementation status remains as described.Swipe to inspect the full diagram.
Wenjie · Hilbertforge
Decide what to investigate
Mission, knowledge, capabilities and needs constrain pursuitworthiness. The proposed initial generators examine contradictions, assumptions, measurement needs, capability changes, social needs and negative results. Coverage certificates and prior-work collisions test opportunities. Discovery preserves alternative solution space rather than choosing an algorithm before seeking an application.
Question certificates and portfolio selection
Question certificates state the target, value, rivals, minimal test and resources. Preflight and blind review precede Pareto and quality-diversity portfolio selection, followed by human retention. Teaching and review subscribe independently and cannot write each other’s judgements. Certificates can feed idea discovery; the Eulerforge handoff name in the original design is a historical interface.
Current delivery and integration role
The local package contains architecture, Question Card schema, event contracts and review rules. Its first vertical slice is described as a proposed implementation. It therefore contributes a design workstream for question value and experimental envelopes rather than a deployed opportunity-discovery service. Integration should preserve the distinction between question selection and mechanism search.
Evidence and criticism change the next research action. Conceptual workflow.Swipe to inspect the full diagram.
Innovation Generator · Ariadne
Explore the next discriminator
Evidence-guided search, substantive revision and testable research dossiers within AutoResearch.
How current results shape later design
Evidence-guided search, substantive revision and testable research dossiers within AutoResearch.
Hand off to experiments and structural recombination
Outputs are candidates with mechanisms, predictions, precedents, actions and states, plus exploration records that can inform experiment contracts or Fusion parents. Acquisition is not a causal experiment, and one discriminating probe does not automatically confirm a mechanism. Host observations and independent review receive these objects; open-domain interfaces still need implementation domain by domain.
Structural inspection and scientific-effect experiments are distinct stages.Swipe to inspect the full diagram.
Lancer Fusion
Find structure, then verify the combination
The engine accepts 2–N questions or ideas, compiles typed structures, searches mappings and Generic Space, and establishes a Fusion Contract. A load-bearing Bridge connects to constrained FusionDSL, followed by type, scope, logic, interface and evidence checks. Problem and idea fusion share a kernel but use different conflict semantics and objectives.
One result with complete lineage
Each stage of the multi-parent tree checks A, B and Bridge necessity. Accepted candidates enter a Pareto/quality-diversity archive; the product returns one selected card. Mappings, failures, repairs and witnesses remain inspectable through CLI, API and local UI. Potential grades, structural levels and evidence maturity are separate, and target grades are goals rather than guarantees.
v0.2.0 provides deterministic structural execution, immutable traces and contract smoke tests. Real prior-art retrieval, domain experiments, trained semantic providers, independent review and sign-off remain external boundaries. Within the programme it recombines candidate mechanisms; bridge and parent ablations, ordinary combinations and matched-budget experiments test substantive gains.
Sources, frozen experiments and evidence states retain their relationships.Swipe to inspect the full diagram.
Research Hub
A shared entry with distinct evidence states
The workbench unifies source records, skill reviews, engineering maturity and research judgements while retaining reading depth, versions, inactive historical sources and originals. Status, search, source, verification and assessment operations expose provenance and benchmark adequacy. A PDF, extracted text or valid hash does not imply completed semantic reading.
Frozen modules and improvement headroom
Runs invoke registered frozen local modules rather than arbitrary shell commands and write new directories preserving existing evidence. Contracts lead into training or policy compilation, refutation, whole-family tests and recalculation. A capacity diagnosis shows the original threshold is unattainable in a fixed policy library, motivating headroom preflight. Execution checks can pass while a scientific effect remains unidentified.
Workspace and platform integration status
A local research process has run, but historical engines, review, Fusion, teaching, clients and distributed services are not one integrated product. Research Hub therefore supplies the evidence and experiment workstream, using integration audits to separate reusable mechanisms, historical checks and unfinished interfaces. Recursive self-improvement requires changed actions, full-cost unseen-task comparisons and independent promotion.
Evidence-constrained manuscript and figure compilation
Writing and figures share author evidence, then meet cross-checking.Swipe to inspect the full diagram.
LSMFC
Register facts before compiling expression
Theory, methods, code mappings, experiments, literature and visual evidence enter shared registries and seven-layer AVIR. Central claims require evidence, warrants, assumptions and potential defeaters. Diagnosis leads to joint story–figure planning, separate manuscript and figure compilation, then checks that modalities support the same claims.
Dependency propagation and the visual firewall
Experimental changes mark related claims, paragraphs, figures and sign-offs stale, returning them to checks. Source and sentence-level provenance travel with artefacts. Visual refinement may change expression but not data, statistics, coordinates or scientific relationships. Research gaps generate upgrade tasks; language providers cannot introduce facts, numbers, citations or broader scope.
From checkable findings to a release bundle
v2.0.0 provides CLI, API, a local workbench, review and human sign-off, plus manuscripts, editable vectors and auditable release bundles. It is a non-training toolchain without dedicated model-training or cluster infrastructure. As the programme’s delivery workstream it receives checked facts and claims; polished writing and demo builds do not establish scientific validity.
An offline role-design scaffold; real-paper validation remains pending.Swipe to inspect the full diagram.
Themis Court
The contribution-review and writing module in AutoResearch, connecting lens routing, independent critique, evidence profiles, and deterministic checks with research taste and author-guided writing.
For author-side self-audit, laboratory discussion, and research education. Current checks use offline fixtures; real-paper First Flight, live retrieval, substantial calibration, and blinded evaluations remain pending.
Discovery, screening and synthesis; metadata checks do not verify every scientific claim.Swipe to inspect the full diagram.
Literature Review
A structured route through source discovery, record processing, staged screening, thematic synthesis, and citation review, supporting evidence preparation within AutoResearch.
This page describes a workflow and its supporting utilities. Source access, selection decisions, and substantive evidence checks remain part of each review.
InkTide brings inspected paper lessons, author evidence, original prose and figure design into a connected writing workflow, with a separate review of the result.
Author evidence and reading lessons have separate inputs; production meets independent checking.Swipe to inspect the full diagram.Explore InkTide →
Current implementation & research scope
This is a research programme and local engineering collection with multiple workstreams. Cross-module identities, permissions, services and meta-policy promotion still require integration and validation. It does not claim a fully deployed platform, RSI certification or method originality.