Skip to content

// A research observatory · Est. 2026

Your unpublished work, on the same map as the literature.

Orion embeds your notes, experiment writeups and PDFs into the same space as the published papers in your field. You see which papers sit closest to work nobody else has read, and you ask across both at once.

// The claim

Both halves of the work, in one space.

A researcher forgets what they wrote eight months ago. They read a paper without noticing it sits next to an experiment they already ran. They propose a direction their own notes already marked as a dead end.

None of the usual tools can catch that, because none of them holds both halves.

Orion chunks and embeds your own material with the same model as the papers, so a melt curve from March and a preprint from last week land as points in the same manifold. The distance between them is something you can look at, search by, and ask about.

// What each tool cannot see

  • Search engines

    A ranked list. No shape of the field.

  • Reference managers

    A library of what you already found.

  • Your notes folder

    Disconnected from the literature entirely.

  • Chat assistants

    Have never seen your unpublished work.

  • 05

    Orion

    Your private corpus and the published one, co-embedded and drawn together.

// 01 · The literature

A field, drawn as a manifold.

Every paper is embedded and clustered into labelled regions. Maps like this already exist; Open Knowledge Maps and Nomic Atlas draw them well. This is the backdrop, not the point.

// How it works

Six steps, from corpus to question.

  1. Ingest

    Papers arrive from arXiv on a schedule. Your notes, writeups and PDFs arrive when you write or upload them. Both are cut into passages and embedded by the same model.

  2. Map

    Everything lands in one space. Papers cluster into labelled regions, and your own passages sit among them, drawn so you can always tell the two apart.

  3. Generate

    Candidate directions are computed geometrically: between two clusters, along a field's trend, or beside a promising point. Only those landing where little published work sits are kept.

  4. Ground

    Each direction is written up citing papers retrieved near it, and a validator rejects any citation outside that set. It checks that a source exists and was retrieved, not that it supports the sentence.

  5. Ask

    Ask one paper, one note, or everything at once. Answers cite the passage, paper or note they came from, and say plainly when the material does not cover the question.

  6. Speak

    When your hands are in a glovebox, ask out loud. Orion answers in its own voice, with captions, and keeps the conversation for later.

// Limits

What Orion is not.

A tool that overstates its own novelty teaches its users to trust it in the wrong places. These are the edges, written down.

  • Not the first map of science.

    Open Knowledge Maps and Nomic Atlas already draw the literature as a zoomable map, and draw it well. What is ours is your unpublished work placed on that map.

  • Not a citation checker.

    The validator confirms that a cited paper was retrieved for the claim. It does not judge whether the paper supports the sentence. Tools like scite do that, and Orion does not.

  • Not a discovery engine.

    A sparse region can mean nobody tried, that people tried and it failed, or that it cannot be done. Orion marks it as a question worth checking, and the first protocol step is always to look for negative results.

  • Not a substitute for the bench.

    Generated directions tend to read better on paper than they perform in the lab. Orion points to passages so you can read them yourself.

Ready

// Voice

Ask out loud.

Voice is a convenience, not the reason to use Orion. It is there for the moment your hands are busy: hold to speak, and the answer comes back in Orion’s own voice, sentence by sentence, with captions and the same citations as typed chat.

// You

Press play to hear a turn.

// Orion

Your batch seven curves show the gain only in phosphate buffer, which points to charge rather than packing. No published paper I can see tests the same variants at two ionic strengths.

Chart your own corpus.

Start with one note. Within a minute it is a point on the map, with the published work nearest to it already drawn in.