Deep search
When your library runs out, the Scout searches open papers and the web, screens what it finds in Markovian steps, and answers from what you import.
Where it starts
The Deep search toggle in the chat box, Find sources on a This isn't in your library yet. answer, a folder's Find sources on this topic, find sources in a space's add tray, the Research starter, a space template, or the orb.
What it does
Plan
The Scout turns the question into a topic (the new space's name), up to three paper searches and
one web search. In this run Nemotron 3.5 Lightning gave no answer within 12 s, so gpt-6-luna
planned it: Sleep Memory Consolidation, with searches like sleep replay synaptic plasticity memory consolidation.
Search
Each paper search asks OpenAlex for open-access articles, preprints and reviews with a PDF, 12 at a
time, skipping DOIs already in your library. The web search uses OpenAI's gpt-5-nano with web
search (up to four pages, bot-check pages dropped), or Wikipedia's own search when that fails.
Screen, in Markovian steps
Nemotron 3 Ultra reads the finds in batches of eight (title and abstract), and only a shortlist of up to eight, with a score and a one-line reason each, carries from one step to the next. This run screened 38 sources in 5 steps. Markovian reports what that saved against one long prompt: 6,569 tokens used, about 12,472 saved (an estimate it labels as such).
You choose, or it imports
Import N sources? lists the shortlist with each reason. Pick, steer it (only human studies, after 2015…), Stop it, or let it import without asking (the Research starter does, and so does Import sources without asking in settings). Papers go into a new space named for the topic, or into the folder you asked from.
Read, then answer
Each import is read like an upload, open PDF first, into Zetaris. Once they're ready (it waits up to 8 minutes), the question is answered from them with Compare on, every quote checked. This run took 4:23 and imported 7.
Open papers only
Deep search reads only papers it can download; a paywalled paper becomes This paper has no free copy online. Upload its PDF to read it here. A failed deep search isn't retried, because its calls are paid.
How Markovian is used, and the rest of Meterless: Meterless.
A real run, step by step
One question across three papers, as the app showed it, and the same question recorded step by step from its trace.
The same check, everywhere
Practice answer keys, draft checks, audio briefs, articles and watch findings are all checked against your sources before you see them.