The operating system for research
Discovery, with people in the loop.
AI now writes, runs and analyses faster than anyone can check. Continual Science is an operating system for research in which every result is checked, understood and answered for by a person.
01 · Why now
AI is eating science.
The length of task an AI system can finish on its own has doubled about every seven months since 2019.
- 2019 · secondsLanguage models finish tasks that take an expert a few seconds.
- 2023 · minutesGPT-4 handles tasks of several minutes.
- 2025 · hoursThe newest models work through tasks that take an expert hours.
- If the trend holdsAI finishes a working day’s task on its own around 2027.
01 · Why now
The frontier labs have moved into science.
Seven launches in ten months. Some are free for academics. All of them run on the vendor’s own cloud.
Claude for Life Sciences
Anthropic · source
$70M seed for Kosmos, an autonomous “AI scientist”
Edison Scientific · source
Prism, a workspace for writing papers
OpenAI · source
Gemini for Science
Google · source
Microsoft Discovery, generally available
Microsoft · source
Claude Science, in beta
Anthropic · source
ChatGPT free for 100,000 academic researchers
OpenAI · source
02 · The gap
Output grows faster than anyone can check it.
- Until recentlyPeople could check what was made: every sample, every run, every sentence.
- NowProducing results gets cheaper every month. Checking them does not.
- The gapEvery result nobody checks is a result nobody understands.
What AI producesWhat people can checkUnchecked: made, but checked by no one
02 · The gap
What a lab loses when nobody checks.
- Understanding
- Results arrive faster than anyone can explain them. Papers cite numbers no one in the lab can reproduce.
- Adaptability
- Skills we stop practising fade. The next unfamiliar problem meets a lab that can only prompt.
- Independence
- Every step runs through one vendor’s model, pricing and terms of use.
- Data
- Unpublished data, samples and ideas leave the institute to be processed elsewhere.
02 · The gap
Science expects a person to answer for it.
A person is accountable; AI cannot be an author.
COPE position statement on authorship and AI tools, 2023 · ICMJE Recommendations
Use of AI is disclosed and documented.
Nature Portfolio editorial policy · DFG statement on generative AI, 2023
Every result can be traced back to its data.
AI text published to inform the public needs a label, unless a person reviewed it and answers for it.
EU AI Act, Article 50, applies from 2 August 2026
03 · Continual Science
An operating system for discovery with people in the loop.
It follows the steps science has always taken, from question to paper, and keeps a named person responsible for each one.
03 · In use
See it in use.
Short loops of a lab at work, each with one person doing one real thing. Each ends with the main point, and you can replay any of them.
01 · Chat
Ask a colleague, get the run back.
Type @ to mention someone and # to link a task. The reply comes back with the run attached.
Alex · PI
Loading the demo…
02 · People
Find who knows about something.
Search the lab by skill. The card shows what someone knows, and you can message them from there.
Leo · PhD student
Loading the demo…
03 · Tasks
Clear a late task and unblock a run.
A late task is holding up a run. Alex assigns the GPU and marks it done, and the rerun is ready.
Alex · PI
Loading the demo…
04 · Calendar
Move a meeting when the room is taken.
The seminar room is double-booked. Sofia drags the lab meeting to 14:00, sees everyone is free, and lets them know.
Sofia · Postdoc
Loading the demo…
05 · Project
Add a task to a deadline.
See what a deadline depends on, then press T to add a task with a person and a due date.
Alex · PI
Loading the demo…
06 · Analysis
Ask a question, check the answer.
The assistant reads the data and suggests steps. Sofia changes one before running, then checks the result.
Sofia · Postdoc
Loading the demo…
07 · Expert analysis
Change a setting and rerun.
Press E to see methods, settings and code. Change a value, rerun from there, and compare with the last run.
Leo · PhD student
Loading the demo…
08 · Map to code
Go from the overview to the code.
The map shows every step. Validate confirms it still runs the same, then zoom in down to the code.
Leo · PhD student
Loading the demo…
09 · Report
Put a live number in a report.
Type # to insert a number from a run. A colleague sees where it comes from and checks the paragraph.
Anna and Alex
Loading the demo…
03 · How it works
Built bottom-up, along the steps of science.
Every step lives on one record. The assistant proposes; a person checks.
- 1 · Question Start with what you want to know. The hypothesis, the plan and who does what live on the record.
- 2 · Data Samples and files arrive with their origin: who made them, when, and with what.
- 3 · Analysis The assistant drafts the steps. A person reads them, reruns them and checks them.
- 4 · Figure Every value in a figure links back to the run that made it.
- 5 · Paper Numbers in the text stay live until the paper is released. Nothing reaches it unchecked.
1Question
Hypothesis, plan, who does what
Checked by Alex
2Data
Samples and files, with their origin
Checked by Jonas
3Analysis
Steps you can read and rerun
Checked by Sofia
4Figure
Every value traced to its run
Checked by Anna
5Paper
Numbers stay live until release
Checked by Alex
Assistant layer: drafts and runs, marked “Not checked”A named person checks each step
03 · How it works
Nothing counts until a person checks it.
Assistant · Not checkedChecked by Alex · Thu 24 Sep, 10:40
1,146 differentially expressed genes
The assistant’s work sits on a dotted layer and says so. When a person has looked at it, their name and the time become part of the record, and the dots wash away.
Illustrative example from a fictional lab.
Not checked until checked
Assistant output keeps its stamp until a named person checks it.
Every number traces back
Each value links to its run, data and code, and stays live until release.
Your data stays with you
Self-hosted or on trusted European hosting, with open models you choose.
One record, one place
Chat, tasks, samples, analysis and writing share one record.
04 · For institutions
Your institute’s AI, on your infrastructure.
Self-hosted or trusted hosting
Run it on your own servers or with a European provider you already trust. Research data does not have to leave.
Open models you choose
Use open models on infrastructure you control. What your lab writes does not train someone else’s model.
Built for good research practice
Named checks, traceable numbers and labelled AI output, designed around the DFG code, GDPR and the EU AI Act.
We are looking for research groups and institutes to pilot Continual Science with us.
Request a pilot05 · How it is set up
Three parts, run where you trust.
People use the app in a browser. A collaboration server keeps the lab’s shared record, and a compute server runs analyses where your data and machines are. Both servers can sit inside your institute’s network.
The app
In the browser, on any computer
macOS, Windows and Linux. Nothing to install.
AMAWJBLS
On campus, or from home through VPN
Inside your network or VPN
Collaboration server
The lab’s shared record
Chat, tasks, samples, analyses, reports and every check, kept in sync for the whole team.
Helmholtz CloudYour serversHosted by us
Compute server
Runs analyses where the data is
On your cluster, grid or GPU servers, for example with Slurm. Raw data does not have to move.
Your HPC or gridGPU serversA single lab machine
AI models
The assistant uses your choice
- BlabladorHelmholtz’s own model service, for Helmholtz members
- Your model serviceOpen models on your own compute, through a standard API
- Our model serviceRun by Continual Science · in development
Everything can run inside one institute. Or keep compute at home and let us host the collaboration server; your raw data still stays where it is.
06 · Coming soon
Preprints you can verify.
A new kind of preprint server. It holds the manuscript together with its data and code, and shows which results in the paper can be verified end to end.
Example · preprint v1 · fictional lab
Colonisation reshapes the colonic epithelium within six hours
Weber, Becker, Rossi, Morgan
- ✓ Re-run
Six hours of colonisation changes the expression of 1,146 genes.
Run #14 · data and code attached · result matches - ✓ Re-run
41 of 200 hallmark hypoxia genes respond.
Run #14, step 5 · result matches - ● Checked
Vegfa rises in qPCR at 6 h.
Data attached, no code · checked by the authors - ○ Text only
This matches earlier reports in conventional mice.
Cites the literature · nothing to re-run
Paper, data and code together
Post the manuscript with the data and code behind each figure and number. Readers can open any value and see where it comes from.
See what is verifiable
Every result says whether it was re-run from data and code, checked by the authors, or is text only. No more guessing.
An API for journals
Journals can ask for a reproducibility report automatically when a manuscript is submitted.
GET /v1/preprints/{id}/reproducibility { "results": 4, "rerun": 2, "checked": 1, "text_only": 1 }
The system for doing research together with AI.
People and AI on one record. The assistant proposes; a person checks.
hello@continual.science