Essarion Science.
An autonomous research scientist for biology and medicine. Ask a scientific question and it plans the study, pulls real records from public biomedical databases, runs the analysis in a live notebook, and adversarially stress-tests every claim — then answers with each step and source attached.
§ 01What it is
Essarion Science does the experiment, not just the reading. Where ResearchAnything.ai reads across the web and returns a cited report, Science takes a scientific question and runs it as a study: it forms a hypothesis, writes an analysis plan before it touches data, retrieves primary records from public biomedical resources, executes real code against them, and then tries to prove itself wrong.
The product surface is science.essarion.com. Underneath it is a biomedical research engine — an internal workbench the codebase calls BioMed — built from four parts:
- A reactive planner loop that decides which tool to run next based on what the run has learned so far, rather than following a fixed script.
- A live notebook kernel whose namespace persists across cells for the length of a run, so analysis builds on itself.
- Source-grounded retrieval across thirteen public biomedical databases, fetching records rather than summaries of records.
- A self-verification pass — preregistration, Bayesian posteriors, multiverse and sensitivity analysis — that runs before anything is reported.
§ 02Why it's different
It computes, it doesn't narrate
A literature agent produces prose about studies. Essarion Science produces prose and the analysis behind it: pooled effect sizes it calculated, sequence work it ran, single-cell QC it performed, protein structures it folded and rendered. Every number in the report traces back to a notebook cell you can open.
Primary records, not abstracts
Retrieval reaches the underlying resources directly — PubMed and Europe PMC for literature, GEO for expression series, ChEMBL for bioactivity, UniProt and Ensembl for sequence and annotation, ClinVar for variant significance, the PDB and AlphaFold for structure. See Data sources for the full roster.
Falsification comes first
The user's claim is treated as H1 — the hypothesis under test — never as a conclusion to defend. Before analysis begins, the plan is written and hashed; deviations after that point are reported as exploratory rather than confirmatory. After analysis, the finding is re-run across the whole space of reasonable analytical choices, and a report is only allowed to claim what the evidence gates permit. This is covered in Verification.
The science stack is key-free
Every connector and skill in the workbench talks to public scientific APIs or runs local compute. There is nothing to configure and no credential to supply for the databases themselves — optional keys only raise rate limits.
§ 03Depth modes
Every run picks a depth mode, which sets the literature pool, the evidence thresholds, the iteration and wall-clock ceilings, and which optional tools are available. deep_dive is the default.
| Mode | Shape | Literature pool | Deep-read | Ceilings |
|---|---|---|---|---|
| scan | A short research brief. No paper is emitted; the posterior is still populated. | 16 candidates | 6 sources | 60 iterations · 10 min |
| deep_dive | A full preregistered paper. The default. | 36 candidates | 18 sources | 200 iterations · 60 min |
| exhaustive | A definitive paper with a replication graph and figures extracted from source PDFs. | 72 candidates | 32 sources | 500 iterations · 4 hr |
Depth also gates tools. scan runs without the pattern-analysis family, the multiverse curve, the sandboxed shell, and PDF vision — there is rarely enough pooled literature in a one-to-two-minute brief for those to say anything. deep_dive and exhaustive enable all of them, and raise the replication thresholds a claim must clear.
§ 04What a run produces
A paper, or an honest brief
A completed deep_dive or exhaustive run produces a structured paper: hypothesis, preregistered plan, methods, results with pooled effects, figures, limitations, and references. If the evidence does not clear the gates — too few effect sizes, too few independent sources, too few distinct labs — the run reports a brief instead. It does not upgrade thin evidence into a paper.
The notebook
Every executed cell is retained with its output, figures, and provenance, so the analysis can be read the way you would read someone else's notebook.
The posterior
A Bayesian posterior on the hypothesis, updated from the live evidence stream, plus the verdict shifts that moved it. The narrative's confidence language is checked against the posterior before the artifact is released.
Exports
Finished work exports to Markdown, LaTeX, BibTeX, DOCX, and PDF.
GET /api/v1/export/{session_id}/markdown
GET /api/v1/export/{session_id}/latex
GET /api/v1/export/{session_id}/bibtex
POST /api/v1/export/{session_id}/docx
POST /api/v1/export/{session_id}/pdf§ 05Where it fits
Three products, one engine. ResearchAnything is the research engine and the flagship; Sourcipedia is the living encyclopedia it writes and keeps current; Essarion Science is the autonomous scientist that goes past reading into analysis. They share the platform's vocabulary — projects, runs, sources, citations — documented in Core concepts.
Use Science when the question is biomedical and the answer depends on data: does this effect hold, how strong is it, does it survive when the analytical choices change. Use ResearchAnything when the question is broader and the deliverable is a defensible cited report.