the research console
An agent that reads the literature and runs the experiments.
You describe the research question. The agent builds a durable, citation-verified corpus of papers and sources, runs tracked experiments beside it, and writes the findings back — while live structured views show you exactly what it did, claim by claim, run by run.
jlit — durable literature projects
A corpus you can audit, not a chat you have to trust.
Discovery, screening, ingest, notes, and a growing wiki — recorded as plain files and SQLite in a git repository. The wiki is the deliverable: ordinary Markdown where every citation is live.
Verbatim evidence, machine-checked
Claims anchor an exact quote to an immutable extraction of the paper. validate re-checks every anchor; a claim that can't point at its evidence is an error, not a style problem.
Screen wide, read deep
Searches register everything they touch under stable ids — metadata is free. Only papers worth reading get ingested: PDF fetched, text extracted, reading notes scaffolded, all provenance kept.
Views that drill to the line
Wiki, corpus screening table, paper inspector, full-text search, evidence graph. Click a citation chip and land on the quoted span, highlighted, inside the extraction it came from.
rx — automated experimentation
Experiments on the record, not in the scrollback.
The agent proposes, runs, and analyzes experiments in a tracked workspace: every run captured with its code, metrics, and logs.
Runs, benchmarks, leaderboards
Each run is an artifact with captured metrics; benchmarks rank submissions on live leaderboards. Nothing is a claim in prose — it's a row you can reopen.
Findings and dead-ends
Distilled knowledge is first-class: findings record what worked, dead-ends record what didn't — struck through in the feed, never silently forgotten, never retried by the next agent.
Worker agents
Spawn parallel workers in isolated worktrees to explore branches of the idea; their runs land on the same leaderboards, their chats sit beside the main one.
composed sessions
One session, both minds.
A literature project can host any number of experiment workspaces. The agent surveys the field, forms a hypothesis, runs the experiment — and the write-up is ingested back into the corpus as a citable source, so the wiki cites your own results next to the papers that motivated them. The loop closes with validate: 0 errors.
deployment
Yours, all the way down.
A machine per person
Signing in wakes your own pre-provisioned isolated appliance — dedicated machine, private network, persistent volume. It sleeps when idle; your projects don't.
Your keys, scoped
Bring your own agent subscription and provider keys. Credentials are stored server-side, injected per-tool at spawn — never broader than the work needs.
Plain files underneath
A project is a directory: Markdown, PDFs, SQLite, git history. Clone it, grep it, take it with you. The console is a view, not a silo.