Loading the page.
Mission
Biomedical knowledge is fragmented across papers, databases, models and ontologies. Catalyst is assembling a provenance-aware representation of human biology that can support mechanistic research now and quantitative simulation over time.
The first layer is a curated evidence graph: the literature becomes computable while preserving the source, study context and uncertainty behind each finding. The strength of the evidence behind a claim is computed from the study it came from — its design, its sample size, whether anyone replicated it, who funded it — and application code has no grant to set it. In vitro evidence is never stronger than very limited, whatever else is true of it. That is not a policy somebody promises to follow; it is a function, and it is written down.
The consequence is worth stating plainly, because it is the whole point: no sale can move an evidence strength. Nothing commercial is an input to the function, and when a compound page does link to a product, it will show the evidence strength the graph computed, very limited included, rather than hide the product until its evidence looks strong. A bar that decides what can be sold is a bar with a reason to move.
A staged technical roadmap: build the evidence substrate, make it useful for mechanistic research, then earn the right to predict with quantitative models.
Connect biological entities, reference knowledge and reported findings through stable identifiers and traceable provenance. The live graph distinguishes sourced findings from reference relationships and mechanistic inference.
The next layer will traverse the graph, assemble evidence across sources and construct mechanistic hypotheses. Those hypotheses will preserve their component provenance and remain visibly distinct from findings any source directly reports.
The long-term system will connect reactions, compartments, parameters and initial conditions to model how interventions propagate through physiological systems, then compare predictions with independent human evidence.
These are the ones a commercial platform finds expensive, which is why they are written into the schema instead of into a policy.
Symptom to cause is guesswork, the corpus holds no priors for it, and this platform’s apparatus would make a wrong guess more persuasive rather than safer.
A trial that found nothing is a finding and is shown with its own evidence strength. A page that only lists what worked is an advertisement.
Where findings conflict, both render, strongest evidence first. There is no net verdict and no “on balance”; the disagreement is the finding.
Bulk third-party data loads into its own table and can never become a finding. The only route into the evidence layer is somebody reading the paper.
Where sleep, training or diet is what the evidence supports, that is what the page says first.
No self-reported certainty is an input to any layer. It is a fact about the model, not about the world.
Built and operated by Gabe Upton, under The Catalyst Project, LLC. Questions, corrections, or a claim that looks wrong: gabe@catalystproject.ai.
How a claim gets from a paper into the graph, and every rule that can stop it.
What this is built on, and the terms each part is used under.
1,385 findings and the entities they connect, filterable by relationship and evidence strength.