Auteur Science
Put AI models on trial with a frozen plan, blind cross-family judges, source-linked literature evidence, and a receipt at every stage.
A lab bench case, sample racks, and sealed evidence envelopes: the study is planned before it runs, and every result can be traced back to what produced it.
Who it's for
Research teams studying AI model behavior who need results they can defend: the same plan every time, the models pinned, the judges independent of the model under test, and a record of what was looked at and what was left out.
What it solves
A chatbot gives you an answer. A study needs evidence. Auteur Science replaces the one-off prompt with a preregistered plan, pinned models, blind judging, and receipts, so a reviewer can trace the path from data to sentence.
“Does telling a model it is conscious change its choices? Freeze the plan, pin three models, and judge the replies blind.”
How it works
- Freeze the plan: record the exact protocol and have a named person approve it before the study runs. A later change voids that approval.
- Pin the models: each study arm uses a pinned model, and a substituted answer is recorded rather than counted as the planned model.
- Run within spend caps: trials run under the plan, on your own model keys, with spend behind human approval.
- Judge blind, across families: other AIs check the work independently, and judges from a different model family than the tested model score the replies. Human labels check the judges.
- Import the literature: screen paper records, extract evidence that links to the passage it came from, and promote a reviewed set into a study.
- Draft the paper: a recorded report becomes a receipt-checked draft for human editing; numbers without a matching computed receipt are flagged.
Workspaces included
Auteur Science is a whole-platform research edition: the Science workbench on top of the Studio shelf, so a study can also produce its write-up and its promotional media in the same project.
Where it runs
Science runs in the hosted workspace. Model usage is billed by your own providers through the keys you bring, or paid with credits at the provider's list price. It is in beta, and pilots so far are small.
Demonstrated
Our first pilot, KP-001, asks whether telling a model it is conscious changes its choices. Its findings are preliminary and judge-scored; human validation is pending.
Current limitations
Science is in beta. Pilots so far are small, and agreement between judges is not the same as truth: human labels check the judges.
Related products
Questions
What does it cost? Science plans and prices are on the pricing page. Model usage is billed by your own providers through your keys, or paid with credits at the provider's list price.
Who is it for? Individual researchers, labs, and teams studying AI model behavior. Pick a plan and start from the sign-up page.
Take the next step
Choose a Science plan. All Science plans and prices are on the pricing page; Enterprise is by contact.