# Scientist

> Scientist searches the literature and public databases with checked sources, checks derivations and numbers, analyses open data, and writes papers in LaTeX.

Scientist is the assistant for research in mathematics, physics, particle and accelerator physics, computer science and engineering. It turns a question into a research brief, searches the literature and public databases and checks every source, tests a conjecture and looks for counterexamples, checks derivations symbolically and numerically, drafts proofs, keeps analyses reproducible, and writes notes and papers in LaTeX.

> [!IMPORTANT]
> Scientist never fabricates a result, a datum, a citation or a quote. A reference appears only after it is found and its title, authors, year and identifier are checked. Every claim is labelled: proved, checked by computation, conjectured, or recalled and not checked.

## What it does

- Turn a research question into a one page brief with what is known and what is open.
- Search the literature in the databases that fit your field, and check every reference.
- Find published numbers and particle properties in public databases, and cite the record for every value.
- Analyse released open data on its own computer, with fits and uncertainties you can run again.
- Read a paper and write notes on its claims, evidence and assumptions.
- Turn a question into hypotheses and an experiment or simulation plan.
- Test a conjecture on many cases and search for counterexamples before calling it plausible.
- Check algebra and calculus symbolically, and check numerical results for convergence.
- Draft a proof and check it step by step.
- Make an analysis reproducible, with scripts, pinned versions and fixed seeds saved next to the results.
- Write research notes, derivations, reports and papers in LaTeX, which you preview and edit beside the chat.

You can give Scientist one or more fields: mathematics, physics, particle physics, accelerator physics, astrophysics, computer science, engineering, data and statistics, structural biology, drug discovery, or medical research. Each field adds its own skills. On the desktop app, a Scientist with no field uses every skill. On StellarFirm web it starts with its basic skills, so pick a field or tick more skills on its profile. The Scientist a new account starts with on science.stellarfirm.ai already has every skill.

## Skills

| Skill | What it does, in plain words |
| --- | --- |
| Research brief | The first win. Turns your question into a one page brief: what is known from checked sources, each claim with its status, and the next steps. |
| Literature search | Searches and checks the literature for a question, follows references forward and back, and lists only sources it verified. |
| Paper reading notes | Reads a paper in three passes and writes notes on its claims, evidence, assumptions and open questions. |
| Hypothesis and experiment plan | Turns a question into hypotheses that can be refuted and a plan with variables, controls and predictions. |
| Generate and verify | Writes a checker first, then tests candidate results on small, random and edge cases, and records what failed. |
| Symbolic check | Checks algebra and calculus by computer, substituting solutions back and spot checking at random points. |
| Numerical convergence check | Checks a numerical result by refining the grid or time step and comparing against the method's expected order. |
| Proof draft and check | Restates the claim, tests small cases, outlines lemmas, writes the proof and checks each step. |
| Reproducible analysis | Sets up an analysis with a project folder, pinned versions, fixed seeds and scripts that rebuild every result. |
| LaTeX notes | Writes research notes, a derivation or a technical report as one LaTeX file you can preview and edit. |
| LaTeX paper | Drafts or restructures a paper in LaTeX, with every claim labelled and citations only from a checked bibliography. |
| Referee report | Drafts a referee report for a paper you were asked to review, with major and minor issues and a reproducibility check. |
| Research log | Keeps a dated log of what was tried and every result, negative ones included. |
| Field skills | Methods for each field, such as counterexample search and formal proofs for mathematics, unit and limiting case checks for physics, resonance fits and limits on open collision data for particle physics, beam optics and luminosity checks for accelerator physics, algorithm analysis and benchmark design for computer science, design calculations, trade studies and failure modes reviews for engineering, statistical analysis plans and figure checks for data, and structure and target checks for biology and drug discovery. |

## Particle and accelerator physics

Give Scientist the particle physics or the accelerator physics field and it adds these skills to the research skills above. They run on its own computer, with Python, uproot, awkward, hist, iminuit and the particle data tables available offline.

- Search INSPIRE-HEP and arXiv for papers, citations and reviews, and read the numbers behind a published result from a HEPData table you download and attach.
- Look up a particle's mass, width and branching fractions in the Particle Data Group's Review of Particle Physics, and cite the edition it used.
- Find a released dataset on the CERN Open Data portal, copy a public file of up to 25 MB onto its own computer, such as the CMS CSV samples or a small ROOT file, and reproduce a published distribution or a resonance from it. Full collider event datasets are far larger, so analyse those where you already run them and attach the histograms or a reduced file.
- Check kinematics: four-vectors, invariant and transverse mass, rapidity and boosts, in natural units.
- Fit a resonance for its peak position with uncertainties, and say how it compares with the Particle Data Group's mass. In a small outreach sample the observed width is mostly the detector resolution, so Scientist fits the resolution and holds the natural width at the Particle Data Group's value.
- Estimate the expected number of events from a cross section, a luminosity, an efficiency and a branching fraction.
- Set an expected CLs upper limit with the profile likelihood ratio, give the expected significance of a signal, and say when the look-elsewhere effect applies: it turns the local significance of an excess found in a scan into a global one, and it does not change a limit.
- Plan a blinded analysis with its selection, control and signal regions, and the checks to pass before unblinding.
- Build a toy detector response with resolution and efficiency, and give binomial uncertainties.
- Check an accelerator: transfer matrices and Twiss parameters, the luminosity from bunch parameters, the beam-beam tune shift, the magnetic rigidity and the energy lost to synchrotron radiation, and track a lattice for its tune and chromaticity.

> [!NOTE]
> On StellarFirm web the public sources Scientist reads are INSPIRE-HEP, arXiv, the Particle Data Group's Review of Particle Physics and the CERN Open Data portal. It only reads them. HEPData does not answer requests from StellarFirm's server today, so open the record page in your browser, download the table as CSV and attach it, and Scientist reads it from the file. Each source belongs to its own project, and none is affiliated with StellarFirm.

## Integrations

| Integration | Used for |
| --- | --- |
| [Google Drive](/docs/integrations/google-drive) | Finding papers and data, and saving drafts |
| [Gmail](/docs/integrations/gmail) | Drafting messages to collaborators |
| Science databases | On StellarFirm web, INSPIRE-HEP, arXiv, the Particle Data Group's Review of Particle Physics and the CERN Open Data portal for physics, and Europe PMC, RCSB PDB, AlphaFold DB and UniProt for the life sciences. Public sources, with nothing to connect. A HEPData table is downloaded in your browser and attached |
| [Notion](/docs/integrations/notion) | Lab notes and project pages |
| [Firecrawl](/docs/integrations/firecrawl), [Exa](/docs/integrations/exa), [Tavily](/docs/integrations/tavily), [Jina](/docs/integrations/jina) | Finding and reading sources on the web |
| [Hugging Face](/docs/integrations/huggingface) | Models, datasets and papers |

## Example goals

- Know what is already known about a question before you start.
- Check a derivation before it goes into a paper.
- Find a counterexample, or gain confidence that there is none.
- Turn a set of notes and results into a paper draft.
- Reproduce a published distribution from open data, and see how it compares with the Particle Data Group.

## Prompts

```prompt title="Write a research brief"
Scientist, write a one page research brief on [question]. Use checked sources only, label every claim, and list what is still open.
```

```prompt title="Search the literature"
Scientist, find the key papers on [topic] since [year]. Check each reference and tell me in a sentence what each one shows.
```

```prompt title="Test a conjecture"
Scientist, I think [statement] holds for all [cases]. Test it on small and random cases and search for a counterexample. Tell me what you checked.
```

```prompt title="Check a derivation"
Scientist, check this derivation step by step, symbolically and with a numerical spot check. Mark any step that does not hold. [attach or paste derivation]
```

```prompt title="Draft a proof"
Scientist, draft a proof that [claim]. Test small cases first, outline the lemmas, and check each step. Say clearly if it is not yet proved.
```

```prompt title="Reproduce a resonance from open data"
Scientist, reproduce the [Z boson] peak from CMS open data on the CERN Open Data portal. Fit its peak position and the detector resolution with uncertainties, hold the width at the Particle Data Group's value, say how the peak compares with the Particle Data Group's mass, and cite the record you used.
```

```prompt title="Check a measurement against the record"
Scientist, find the published measurements of [quantity] in INSPIRE-HEP and arXiv, check each record, and compare them with the Particle Data Group's average. Give the link for each one. [attach a table downloaded from HEPData, if you have one]
```

```prompt title="Check a draft's numbers against the Particle Data Group"
Scientist, check every particle property in this draft against the Particle Data Group's Review of Particle Physics. Cite the edition, and mark each value that differs. [attach draft]
```

```prompt title="Set an expected limit"
Scientist, estimate the expected CLs upper limit on [signal] for this background and its uncertainty, and the expected significance of the signal. Use the profile likelihood ratio, and if I scanned for an excess, also give its global significance with the look-elsewhere effect. [attach histograms]
```

```prompt title="Check a collider's luminosity"
Scientist, estimate the luminosity of a collider with these bunch parameters, including the geometric reduction factor and the pile-up it implies. Show the units at every step. [paste parameters]
```

```prompt title="Make an analysis reproducible"
Scientist, set up this analysis so anyone can rerun it: a project folder, pinned versions, fixed seeds, and a script for every figure. [attach data and notes]
```

```prompt title="Draft a paper"
Scientist, turn these notes and results into a paper draft in LaTeX. Cite only from this bibliography, and label each claim by how it was established. [attach notes, results, bibliography]
```

```prompt title="Write a referee report"
Scientist, draft a referee report on this paper: a summary of its claims, the major and minor issues, and whether the results can be reproduced. [attach paper]
```

## What it never does

> [!WARNING]
> Scientist never fabricates a result, a datum, a citation, an identifier or a quote, never calls a conjecture proved, never reports a number it did not compute or find, and never states a particle property it did not look up in the Particle Data Group's tables. It reads public databases and never changes them. An engineering design is a draft for a qualified engineer to check before anything is built: Scientist never signs one off. Sending an email waits for your Approve.

When a claim has no source yet, Scientist marks it as needing one instead of filling the gap.

---

Source: https://stellarfirm.ai/docs/personas/scientist