Computational Co-Science
How AI changes hypothesis generation, experimental planning, coding, modeling, and interpretation across scientific workflows.
RESEARCH & METHODS
Building the Methods for AI-Assisted Science
Santaros Labs is a nonprofit computational research lab studying how scientific teams can use capable AI systems while preserving rigor, traceability, privacy, and human accountability.
OUR MISSION
Expand Scientific Capability Without Weakening Scientific Standards
AI can accelerate literature review, coding, analysis, and scientific writing. Speed alone is not the objective. We investigate when that acceleration produces better reasoning, when it hides error, and what technical and organizational safeguards help research teams distinguish the two.
Plans are not findings
Every Santaros study states its status. Protocols in development and preregistration are never presented as completed evidence.
Four Research Pillars
Reproducibility & Provenance
How to preserve the chain from source data and literature to transformations, code, results, and claims.
Evidence Integrity
How AI handles citations, conflicting evidence, uncertainty, missing data, and requests that exceed the evidence.
Research Team Systems
How shared AI workspaces affect handoffs, review, coordination, expertise, and accountability in scientific groups.
PILLAR 01
Computational Co-Science
Measure the quality of partnership, not just the speed of output.
We study AI as a collaborator across the research lifecycle: question formation, experimental design, simulation, code generation, statistical analysis, and interpretation.
Comparisons separate productivity from epistemic quality. Faster work is useful only when claims remain calibrated, assumptions stay visible, and domain experts can identify where reasoning failed.
PILLAR 02
Reproducibility & Provenance
Make the path from raw evidence to published claim inspectable.
AI-assisted work can involve dozens of transformations that disappear from the final paper: retrieved sources, data cleaning, generated code, model choices, rejected analyses, and rewritten interpretations.
We investigate practical provenance systems that capture enough context for review without creating documentation overhead that scientists cannot sustain.
PILLAR 03
Evidence Integrity
Test whether the claim is supported, not merely well written.
Scientific synthesis becomes dangerous when fluent summaries flatten disagreement, invent citation support, or hide uncertainty. We design adversarial evidence sets that require contradiction detection, source discrimination, and calibrated refusal.
Evaluation combines citation-level checks, expert scoring, confidence calibration, and structured error taxonomies. We report where a system fails as carefully as where it succeeds.
PILLAR 04
Research Team Systems
Study the lab as a coordinated human–AI system.
Research is collaborative. We examine how AI changes the distribution of expertise, review burden, mentoring, handoffs, authorship, and responsibility across principal investigators, research staff, postdocs, and students.
The aim is not to replace scientific training. It is to design shared work practices that make assumptions visible, preserve productive disagreement, and help experts focus attention where it matters most.
We pay particular attention to early-career researchers, whose learning can be strengthened by transparent assistance or weakened by opaque delegation.
How We Work
Each study pairs scientific ambition with explicit controls for validity, ethics, security, and reproducibility.
Preregister
Specify questions, measures, exclusions, and analysis decisions before outcomes are known whenever the design permits.
Work with domain experts
Evaluation criteria come from the science being studied, not from generic impressions of output quality.
Protect research data
Use data-minimization, access controls, and study-specific security practices appropriate to the sensitivity of the work.
Preserve provenance
Record sources, transformations, code, model context, and material human decisions across the workflow.
Replicate
Separate exploratory discovery from confirmatory analysis and include independent reproduction where feasible.
Report limitations
Publish boundaries, null results, and failure modes so future teams know where evidence does and does not apply.
PARTNER WITH US
Study the Future of Science Without Guessing
Bring a research question, an evaluation problem, or a reproducibility challenge. We will start with the study design—not the marketing claim.
Proposed collaborations proceed only after scope, governance, ethics, and data terms are clear.