Research Desk

What it is

One view per research project: the argument, how well it is backed, and what to do next.
Every record is a Markdown note in your vault; the view reads and writes those notes.
Open it with /research or the Open research desk command.
External MCP clients can read and write the same records through the optional MCP bridge.

Fix first appears when records have broken references, unreadable notes, or unverifiable sources; each row opens the note.
Ask Claude lists up to three suggested steps and a box for typed instructions.
Argument shows one card per claim with its status, supporting and challenging counts, and passages.
Document shows the outline or draft with sections drafted, sections changed, and buttons to open it or work on its sections.
Add sources has Link or file, Search papers, and Pull passages.
Unused passages, unread sources, and rejected claims sit in collapsed groups under the cards.

Suggested steps come from the project's notes and need no model request to appear.
Typed instructions and chat steps run in agent mode with the project attached.
New evidence and claims arrive as proposed; you check them before they count.
Draft steps preview a section and wait for Accept before writing to the document.

The first matching row applies.

Status	Shows when
Rejected	The claim is rejected
Needs check	The claim is still proposed
Unsupported	No trusted evidence supports it
Thin	One trusted passage supports it
Challenged	Challenging evidence or a contradiction exists and no limitation is recorded
Changed since drafted	The claim or its evidence changed after its section was drafted
Drafted	Its section is drafted by a model and unedited since accept
Not drafted	It is in the outline and has no draft
Not in outline	An outline exists and the claim is not in it
Ready	No outline exists yet

Audit and analysis findings appear as Fix first rows and as notes on the claim cards.

Only reviewed evidence with a locator and a current source fingerprint counts as trusted support.
Proposed evidence stays visible and does not satisfy the audit.
Review changes evidence records only and ends in reviewed or rejected.
Evidence is stale when the source's current fingerprint differs from the one captured with it; the excerpt needs re-verifying.

Each section is a ## heading named after its claim.
A claude-provenance block at the end of the note records each section's claim, passages, and citations.
The block renders as References in Reading view and Live Preview.
Renaming a managed heading detaches that section until the heading is restored or the outline is rebuilt.
Clean up format converts older notes to this layout.

Revisions carry the grounded section packet and an explicit intent.
The model's response is validated before preview.
You review the proposed result before it replaces the section.
Unsupported citations, silent claim loss, stale grounding, and malformed responses are rejected instead of written.

Searches OpenAlex, enriched with Crossref and arXiv metadata, ranked locally with an optional model rerank.
Results import into the project as sources.
Network requests fire only on explicit actions, and results cache locally.
Enable it under Settings > Companion for Claude > Scholarly discovery; OpenAlex asks for a contact email.
A Zotero user id, plus an API key for private libraries, resolves zotero_key imports to title, authors, date, publication, DOI, url, and abstract.
A failed Zotero lookup still imports the key.