Science lab

StellarFirm for scientists puts one AI Agent Assistant first: the Scientist. It works in mathematics, physics, computer science, engineering, data and statistics, structural biology, and drug discovery. It searches the literature with checked sources, plans experiments, checks results symbolically and numerically, and writes papers in LaTeX. You are the CEO: the lead researcher. You approve every email, calendar hold, and file that leaves the chat.

Who this is for#

Your assistants#

Hire a Scientist per field, for example a mathematician and a data and statistics scientist, each with its own focus.

Suggested house rules#

House rules
1. Never invent a result, a datum, a citation, or a quote. Say what you could not check.
2. Every number is computed or found, never guessed, and every analysis can be run again.
3. A conjecture is a conjecture until it is proved.
4. The Scientist never signs off an engineering design. A qualified engineer does.
5. Emails, calendar holds, and files saved to our drive wait for my Approve.

Preflight#

• Sign in at https://stellarfirm.ai/app with your invitation.
• Connect your files and notes on Integrations: Google Drive, OneDrive, or SharePoint, and Notion if your group keeps notes there.
• Connect research sources: Hugging Face for models and datasets, and the web research tools such as Exa and Tavily.
• Put one folder per project in your drive: papers, data, and figures.
• Write your house rules in the company brief.

Launch: Day 1#

1. Write the research brief.

Scientist, write a research brief on [question]: what is known, the key papers with their sources, the open problems, and what we could test first.

2. Check a result.

Scientist, check this derivation step by step, symbolically and numerically. Mark every step you could not verify. [attach or paste derivation]

3. Approve what goes out. Anything sent or saved waits in Approvals.

First week#

Scientist, find the most cited and the most recent work on [topic]. Give each source with a link, one line on what it shows, and say which you could not check.

Scientist, propose a hypothesis on [question] and an experiment to test it: what we measure, the controls, the sample size, and how we analyse the result.

Scientist, analyse [dataset] to answer [question]. Keep the analysis reproducible, show the figures, and give the uncertainty on every number.

Scientist, draft a paper in LaTeX from our notes in [folder]: abstract, introduction, methods, results, and discussion, with every citation checked.

Scientist, write a referee report on this paper: its claims, whether the evidence supports them, and the questions I should ask. [attach paper]

Steady orbit#

• Every week. Ask for a literature update on your topics, as a routine.
• Every experiment. Ask for the plan before, and the reproducible analysis after.
• Every paper. Ask for a check of every citation and every number before you submit.

The rest of the team#

The Scientist leads here, and the other assistants are there when you need them:

• Teacher: turns a result into a lecture or a set of exercises for students.
• Coder: turns an analysis into production code or a tool for the group.
• Copywriter: helps you edit grant text and papers in your own words.

Pitfalls#

• Trusting a number nobody computed. Ask the Scientist where each number came from.
• Citations from memory. Every source needs a link the Scientist checked.
• Skipping the plan. Agree the experiment plan before you run it, not after.