What Litmus Note does, and what it cannot do

Is this idea worth pursuing? The tool answers it in pieces: can it be done, has it been done, where is the gap, and what would make it different. You make the call.

1. Your idea. You write it in your own words. The tool can sharpen it into a statement, marked as an AI draft you can edit.
2. Search queries. The model drafts phrases for finding existing work. You edit, add, or delete them. Every query and its request URL is recorded.
3. Retrieval. Each query runs against OpenAlex and Semantic Scholar. Records are merged when the DOI matches, or when title, year, and authors overlap. Result counts and failures are recorded per search.
4. Screening. You include or exclude each record and give a reason. AI can sort first, but a sorted record is only decided once you confirm it. Pre-sorting is not screening.
5. Grounded extraction. For each included record, the model fills in method, data and context, headline findings, stated limitations, relevance, and a gap signal. Every field carries a quote from the abstract.
6. Positioning. Themes, what is well studied, and candidate gaps. Each item is tied to a quote and to the record it came from.
7. Overlap. It compares each included record with your idea on five points: problem, population, context, method, outcome.
8. Feasibility. A checklist of what the work needs. Data access, methods you can run, time, funding, team, equipment, ethics. You mark each one met, not met, or unknown.
9. Novelty angles. Changes the corpus suggests you could make, each with a minimum-study sketch. You can read the sketch in the note or download it on its own. These are suggestions, not proof of novelty.
10. Questions, objectives and risks. AI drafts candidates for you to edit. Risks are drafted from the run record.
11. Your decision. Go, revise, or no-go, with your reasoning. The tool never writes this part.
12. Export. The note as Markdown, a JSON archive, BibTeX for the included records, and an AI-use disclosure built from the run record.

Every extracted field carries a quote from the source abstract, and your browser checks that the quote is really there. If it is not, the field is marked unverified and left out of the note. If the abstract says nothing about the method or the limitations, the field reads “not reported in the abstract” instead of a guess. That answer shows up often, and it is the honest one.

The same rule covers refusals. If an abstract states no limitations, the limitations field stays empty instead of guessing. AI-written passages are labelled as drafts that need your revision.

• Not a systematic review. One reviewer, two databases, no registered protocol.
• Not proof that your idea is new. No search of two databases can show that nobody has done this.
• Not a proposal. No budget, timeline, or methods design.
• Not full-text analysis. Abstracts and metadata only, and the note says so.
• Not an author and not a reference manager. AI text is marked as draft, and no AI can take responsibility for the work.

• Publication bias. Published work over-reports positive results. Null results, failed replications, and grey literature are under-represented. A gap you see may be a gap in what gets published, not a gap in what has been done.
• Coverage and language bias. Indexing leans English and Western. Work published regionally, in other languages, or outside indexed journals may not appear at all.
• Two databases, searched non-exhaustively. OpenAlex and Semantic Scholar. Field-specific or regional databases may hold work this search never sees.
• Abstracts only. Abstracts often leave out methods, sample sizes, and limitations, so extraction is thinner than full text would allow.
• One reviewer by default. With a single screener, screening mistakes go unmeasured. The note records what was done, so a reader can judge it.
• Language models make mistakes. Extraction and framing are drafts and can be wrong. Every field shows its quote so you can check it before you use it.
• A snapshot, not a watch. The search is a moment in time, recorded with its date. The literature keeps moving.
• Retraction flags are partial. Integrity flags come from OpenAlex only.
• Disclosure is your job. Publishers require you to disclose AI assistance. The tool drafts a statement from the run record, and you decide what is accurate.

Your idea, searches, screening decisions, and note stay in this browser. There is no account and no server-side copy of your work.

The model stages send your research idea and public paper metadata or abstracts to the model provider you configured, and only after you agree. You can refuse: every stage also has a manual path. Searches go to OpenAlex and Semantic Scholar. Clearing site data deletes everything the tool holds.