Research scope
Autonomous Research supports narrative reviews, critical reviews, conceptual papers, argumentative papers, systematic reviews, and scoping reviews. Work is based on existing literature and conceptual reasoning, under the responsibility of one human author. It does not run experiments, recruit participants, collect new empirical data, or perform meta-analysis.
Different paper types require different reasoning. A conceptual paper develops constructs and relationships; an argumentative paper examines premises and counterarguments. Systematic and scoping reviews require defined eligibility, documented screening, and an appropriate synthesis. The app includes reporting aids linked to PRISMA 2020 and PRISMA-ScR. Completing those aids does not establish methodological validity or automatic compliance.
Questions and feasibility
Discovery evaluates retrieved literature in batches and follows candidate questions with searches for closest existing work and counterevidence. Candidates describe a possible contribution, feasible literature-based method, evidence availability, and limitations. Feasibility labels are qualitative assessments. They are not probabilities of novelty or promises that no similar work exists.
Database coverage, query choices, failed requests, and inaccessible texts can leave gaps in that assessment. When no candidate is feasible, recovery options let you refine the topic, seek additional evidence, or explore narrower questions. You choose which eligible question to pursue.
Plans and sources
A saved plan describes proposed queries, eligibility criteria, extraction fields, and the synthesis approach. Two independent optional checkpoints let you review the plan before searching, and review sources and the outline before drafting. Both default off. Source inclusion and exclusion decisions can be overridden with a recorded reason; AI recommendations remain separately recorded.
Search and research records retain queries, dates, decisions, counts, and access limitations. Long documents are assessed in source-linked chunks with page or section references where available. Retrieved full text is distinct from text actually assessed. A document can have full-text access while its assessment covers only some chunks. Partial-text and abstract-only coverage remain visible.
Contradictory findings and negative evidence belong in the synthesis. A source’s availability or inclusion does not establish that every claim in it is reliable, that its methods are sound, or that its findings apply to the proposed question. The author must inspect important claims and their context.
Three kinds of review
Quotation matching
The checker compares direct quotations with cited source text, allowing supported normalization and marking editorial alterations for review. A matching string confirms textual correspondence; it does not establish that the surrounding argument is supported. An ambiguous quotation parser result or an unavailable source is a review concern, distinct from a mismatch confirmed against sufficiently complete source text.
AI-assessed claim support
The model compares manuscript passages with their cited evidence and recorded execution facts. It can flag contradiction, scope mismatch, overstated certainty, or missing support. These are model assessments that can be mistaken or incomplete. A proposed conceptual argument is distinguished from a claim about external facts.
Human review
You inspect the manuscript, sources, interpretations, and proposed revisions. You decide whether an argument is defensible and whether a concern is resolved. The software’s checks and your proofreading are separate records; neither is represented as journal peer review or independent second-human screening.
Issues and approval
Issues link a manuscript passage to relevant evidence and retain resolution history. Unknown citation identifiers, confirmed direct-quotation mismatches, demonstrably false execution facts, and unresolved human-confirmed errors prevent proofreading approval and downstream journal actions. Deterministic errors require correction. Model-suspected concerns and parser-review concerns may be dismissed with a recorded explanation after inspection.
Editing and export remain available while issues are unresolved. Revisions are proposed as accept/reject comparisons. Accepting revised text creates a manuscript version and invalidates affected checks and approvals. Assessments are tied to manuscript and evidence inputs; current quality checks are required before a new approval. These readiness and ownership requirements are enforced on the server.
Changing writing-style settings alone does not rewrite an existing manuscript or invalidate its approval. Applying a revision does. Methods facts and model attribution in a generated paper are derived from execution records rather than requested as invented prose. Review status is included in the exported research record.
Human decisions
You must select the research gap, proofread the manuscript, provide accurate author information and declarations, and approve each final submission package. Optional checkpoints add earlier review opportunities. The author remains responsible for source interpretation, originality, disciplinary standards, disclosures, and the accuracy of the final paper.
A feasible question is a starting assessment, and a completed draft still requires scrutiny. The tool cannot guarantee a novel contribution, exhaustive searching, scholarly validity, or acceptance by a journal.
Models and records
Research tasks default to NVIDIA Nemotron 3 Ultra (free) through OpenRouter, with no paid-model fallback. The configured research model identifier on this deployment is nvidia/nemotron-3-ultra-550b-a55b:free. Individual research records preserve the model identifiers reported for actual requests. Rate limits and failures can interrupt work; saved checkpoints support resuming work against matching inputs.
The Nemotron free route accepts text only. JSON output is requested in the prompt and checked locally rather than enforced by the provider. Its free endpoint terms describe logging for product improvement and restrictions on confidential and personal information. New projects on this deployment start in public-research mode without a separate confirmation checkbox. Use only public, non-confidential content without personal information. Older projects created outside this mode require review and confirmation before AI work resumes. The app does not automatically detect or redact private information. AI journal-portal and receipt tasks are disabled because they can contain author details. Export the manuscript and complete submission directly with the journal.
Projects retain versions, source decisions, assessments, and execution information. Private project records are not published as examples. The curated sample paper was prepared separately with AI assistance and has its own public curation record; it does not represent a completed research run or a human-approved manuscript. See Privacy for information about storage and external processing.
Publishing
The external-journal route checks scope, article types, editorial practices, AI policies, and fees against official information. Only verified no-fee journals qualify. Documented acceptance rates can inform prioritization among eligible journals; they are historical information, not a prediction for your paper. You approve each submission, and the journal decides whether to accept, request revisions, or reject it.
Autonomous Research Publishing is a planned in-house family of discipline-specific journals. Its human editorial service and submissions are not open. Initial human feedback within one business day is a service target for that planned offering, not a current guarantee. Feedback, acceptance, and publication are separate stages.
Evaluation status
Automated application checks cover workflows such as source access, versioning, issue review, quotation parsing, approval safeguards, and browser layouts. Those checks establish tested software behavior in the stated scenarios. They do not establish the scholarly quality of every generated paper.
The broader comparison covering 18 cases across six paper types and three disciplinary groups, including independent human claim assessment and manuscript scoring, remains incomplete. Research-quality accuracy thresholds and publication outcomes have not been established by that evaluation. The public sample’s curation record is a bounded editorial record, not a substitute for the evaluation.