Evidence before confidence
Connect a claim to its source. Keep contradictory findings and missing information visible.
Research judgments
should leave a record.
Predome is an early-stage project exploring how specialized AI Agents can help organize biomedical evidence, develop research questions and make scientific forecasts that can later be checked.
A conclusion is easier to trust
when you can inspect the work.
Scientific decisions depend on what was known, which assumptions were made and which uncertainties remain. A polished report alone does not preserve those distinctions.
We are building toward a research record that connects the question, source material, Agent responsibilities, calculations, critiques and revisions. A useful result can be a supported observation, a rejected hypothesis or a clearly stated reason to wait.
WORKING PRINCIPLESMake the boundaries
Make the boundaries
as clear as the ambition.
A proposal is a proposal
Research plans and model hypotheses stay distinct from experiments that were actually performed.
Keep the original judgment
Record inputs, parameters and revisions. A later interpretation should not overwrite an earlier claim.
Let outcomes do the checking
Define how progress will be judged. Agreement between models does not establish scientific truth.
INITIAL SCIENTIFIC CONTEXTGastric and
Gastric and
colorectal cancer research.
Our initial questions sit around new-drug development and translational research: mechanisms, relevant populations, clinical design and the interpretation of trial outcomes.
Molecular reference work, host–microbe data and clinical question development provide three complementary starting points. The public cases describe the source, the work performed and the limits of each.
Read the case recordsA defined
research question
research question
Molecular contextHost–microbe contextClinical context
COMPUTATION · EVIDENCE · REVIEW
WHERE THE PROJECT IS TODAY
A focused beginning.
Predome is being developed on the scale of a small team. The present foundation is computational: public-source collection, numeric checks, descriptive analysis, stored research records and a public forecasting prototype.
A model-assisted drafting interface is implemented, but successful live inference has not been verified. Independent cross-model forecasting and dedicated prospective adjudication remain in development.
We do not operate a wet laboratory or claim drug discovery, clinical validation or a proven forecasting track record. Future experiments or collaborations will be described with their actual methods, contributors and results.
Read the current scope