Transparent
workflows
A traceable record of inputs, method choices, computations, checks, and revisions should make an investigation open to inspection. Gaps in that record should remain visible.
A research direction within AI for Science
AI for Empirical Scientific Computing
AI agents for
complex scientific
investigations.
We’re building agents for computationally demanding scientific work. They plan and adapt investigations over many steps, use domain knowledge to select methods, and connect each finding to its data, computation, and validation record.
An evolving research program.
Synthetic illustration · prewritten decisions, calculated example values.
A SUSTAINED SCIENTIFIC INVESTIGATIONSynthetic example
SCENARIO CONTEXTDoes the observed difference
Does the observed difference
support a persistent change?
- Available
- Two measurement groups, with a shared unit.
- Constraint
- Collection conditions still need assessment.
Plan
Compute
Assess
Refine and continue with the evidence retained
Initial plan retainedRevised plan proposed
ONE REPORTED VALUEDifference in group means
Inspect the source observations, formula, assumptions and checks.
Agent reasoningUnderstand context. Choose and revise methods.
Scientific computationCalculate values. Keep their basis inspectable.
Scientific contextQuestions, data and constraints guide the plan.
Long investigationsProgress and evidence carry across stages.
Traceable valuesInputs, methods and checks stay with the finding.
01 WHY EMPIRICAL
An answer needs
a record of how.
Measurements come with conditions.
Methods come with assumptions.
A result becomes useful when we can inspect the connection between the question, the data, the computation, and the checks that support its use.
Within AI for Science, our focus is empirical scientific computing: using observations and measurements to guide computational investigations, with assumptions, uncertainty, and unresolved questions kept visible.
02 RESEARCH PRINCIPLES
Autonomy, with a record
you can inspect.
Our long-term goal is for agents to identify what needs to be established, choose or compose an analytical workflow, and revise the investigation as evidence changes—within explicit permissions and review boundaries.
Transparent
Engineer-directed
control
People should be able to define objectives and execution boundaries, review consequential choices, and decide when a finding is ready to use.
Grounded in
domain expertise
Scientific methods, assumptions, applicability conditions, and professional practices should inform both planning and interpretation.
Explicit verification
and validation
Checks should address both whether a computation meets its requirements and whether its results are suitable for the intended question. Uncertainty and limitations belong with the finding.
03 AN INVESTIGATION
From a question to evidence
you can examine.
Follow one scientific scenario across planning, calculation, assessment and revision. The question stays in view; the record grows as the investigation continues.
ONE SCENARIO · A CONTINUING INVESTIGATIONResearch direction · synthetic example
THE SCIENTIFIC QUESTION
Does the observed difference support a persistent change?
DataSupplied synthetic measurements
AssumptionsCollection conditions remain unconfirmed
ScopeDescribe observations; retain scientific limits
WORK + COMPUTATIONAL BASISQuestion + context
CONTEXTThe scientific question, data and constraints
INVESTIGATEPlan, compute, assess and revise
RETAIN THE BASISFindings keep their sources and unresolved questions
Each value opens its source data, method and checks. The uncertainty estimate retains its assumptions.
Retained basisUnresolved or proposed
AN ITERATIVE INVESTIGATION
The context stays.
The record grows.
Frame the question, choose methods, compute, assess and revise. Keep the evidence and limits with the finding.
CONTINUING RECORD
- Original question and context
- Plans and reasons for revision
- Computations, checks and unresolved conditions
Explore at your own pace.
A compact illustration of a longer investigation. Plans and assessments are prewritten; reported values are calculated from synthetic data. Proposed follow-up measurements have not been performed.
04 INDUSTRY & RESEARCH COLLABORATION
Let’s examine a scientific
workflow together.
We’re looking to exchange ideas with scientists, engineers, and builders of scientific software.
Read our writing, share a perspective, or start a topic in the community. Articles and discussions have their own pages, so useful ideas have a place to live and a link to share.
- What evidence does a question require?
- What should remain traceable as a workflow changes?
- Where is expert review essential?
Visit the community
Articles & topics · Read freely. Confirm to contribute.
Contributions are public. Share ideas and general workflows, with permission to share any material you include.