Request access

Private preview · Windows research IDE · AI research agent

An AI research agent that keeps its evidence

scime ra takes a research idea through literature, falsifiable hypotheses and experiments on your own GPU, then writes a paper draft with source you can replay. Every trial, seed, failure and figure value stays on record, and you approve each step.

Request preview access See how it works

A desktop IDE built on Code - OSS, with a separate research engine. Your code, data and results stay on your machines.

study ledgerexample
  1. Idea and protocol approvedv2 · you
  2. Literature checked against Crossref and arXiv24 refs
  3. 3 falsifiable hypotheses reviewedapproved
  4. Baseline reproducedmatched
  5. H2 screening: no gain on validationkept
  6. H1 multi-seed runsGPU 0–3
  7. Ablationqueued
  8. Selection lock → final testlocked
  9. Evidence report, figures, paper—
snapshot sha256:…seeds 5raw logs ✓failures kept ✓

The research loop

Each stage produces an artifact you can read, edit and approve before the next one starts. Nothing moves forward on an unreviewed protocol.

  1. Idea and protocol

    Start from your own question, or let an autonomous session propose one falsifiable question at a time.

  2. Literature

    Related work from Crossref and arXiv, with identifiers checked against the authority records.

  3. Hypotheses

    Three hypotheses, each with a mechanism, nearest prior work, falsification conditions and planned ablations, then a scientific review.

  4. Code and data

    Datasets, environment and implementation with bounded repair. A smoke check confirms it runs and never counts as evidence.

  5. Execution readiness

    A fingerprint binds code, data, seeds and metrics to the reviewed protocol before any real run.

  6. Experiments

    Baseline reproduction, validation screening, multi-seed runs and ablations, with an optional small-data pilot first.

  7. Locked final test

    Model selection is locked before the test set is touched, and test scores never flow back into proposals or reviews.

  8. Report and paper

    An evidence report, figures and a LaTeX paper draft, delivered as a PDF plus a source archive.

Built so you can check the result

An AI that runs experiments is only useful if a reviewer, or you six months later, can see exactly what happened.

Replayable papers

Every export ships a source archive with reproduce.py and a hash manifest that replays the experiments without any model calls.

Figures from the ledger only

Each plotted value traces to its experiment, seed and code hash. Invented values or changed claims are rejected before a figure renders.

Every trial kept

Source snapshots, hypotheses, reviews, seeds, raw logs, metrics and an event ledger per study, failures and negative results included.

Citations checked

Title, authors, year and venue are matched against bibliographic records, and quoted support must appear in the cited source.

You approve each step

Stage artifacts are editable and versioned. Approval is bound to the exact version, and auto-approval never skips scientific checks.

A paper library

Finished studies are stored as immutable, hash-verified snapshots of the PDF and its source.

Your compute, your models

Compute and model access are separate settings. scime ra brings no GPUs of its own: it schedules work on the machines you already use.

  • Local or SSH Linux GPUs. One independent experiment per GPU, with device leases so jobs never collide.
  • Claude through your own subscription via the official Claude Code CLI. When a usage limit is reached, work pauses and resumes after the reset.
  • Any OpenAI-compatible API with your own key, stored in the editor's secret storage.
  • Self-hosted open models served over SSH with Ollama, bound to loopback with no open port, sharing GPU leases with experiments.
  • Autonomous sessions with limits. By default at most 8 hours, 10 studies and 3 consecutive failures; sessions are resumable.

Honest status

A working prototype, not a finished AI scientist

  • scime ra is a working research-loop prototype and IDE, available as an internal-preview Windows installer.
  • Long unattended (overnight) scientific throughput has not been demonstrated yet.
  • We make no claim of beating other AI scientist systems; there is no head-to-head comparison yet.
  • An exported paper is a research draft. It does not certify novelty, state of the art or conference readiness.
  • Citation checks confirm records and quoted text, not that a source fully supports a claim.
  • Experiments run your code on your machines; the worker is not a security sandbox.

Part of scime

One account across the research loop, from finding where an idea fits to submitting the paper.

Live · free

scime atlas

See where your idea lands among 108,967 accepted ML, vision and AI papers before you run anything.

Private preview

scime ra

Turn the idea into evidence: hypotheses, experiments on your GPUs and a replayable paper draft.

Live

scime draft

Finish the paper with co-authors, check it before you submit and answer the reviews.

Request preview access

We are onboarding a small number of ML and computer vision researchers. Write to us and tell us a little about your work.

Email contact@latenticlabs.com

Helpful to include

  1. Your field and the kind of experiments you run
  2. The GPUs you can use (local Windows PC, or Linux over SSH)
  3. Your model access: a Claude subscription, an API key or self-hosted models

Requirements: Windows 10/11 x64 for the desktop app.