Scientific method
How the research engine decides what is true enough to say
Everything below is implemented as code and applied the same way every time. When the method changes, the version number changes.
Evidence, not assertions
- Verified sources only. A paper counts only after its identity is verified; retracted or unverifiable papers are excluded with a recorded reason.
- Quote-verified extraction. The AI model that reads abstracts must quote them word for word; code discards anything it cannot find in the text.
- Competing hypotheses. Every question keeps several explanations side by side, including “no effect” and “depends on conditions”.
- Independence. Evidence is counted per paper, not per claim, so one paper cannot outvote others.
- Bayesian belief updates (belief@1) with bootstrap uncertainty intervals, recorded with every input so each update can be audited.
- Failure Memory. Rejected hypotheses and negative results are kept and must be consulted before proposing new mechanisms.
Simulations and experiments
- Reproducible models. Every simulation records a manifest (model version, parameters, random seed, code) so it can be re-run exactly; some are re-run automatically to check.
- Uncertainty is part of the result. Signalling models are run hundreds of times with uncertain parameters varied together; results report how many variants agree.
- Pre-registered experiments. A yeast experiment's analysis and decision thresholds are frozen before any data exist.
- Safety first. A rules-first safety classifier labels every proposed physical experiment GREEN / YELLOW / RED. Only GREEN is ever suggested for a home or teaching lab.
Findings: the publication bar
A result becomes a public finding only if the leading explanation reaches at least 70% belief, supported by at least two independent papers, and a human reviewer approves the exact public wording. The bar is never lowered to produce findings.
- Computational predictionA computer model or analysis suggests it.
- Literature supportIndependent published studies support it.
- Yeast validationA pre-registered yeast experiment agreed.
- Professional mammalian validationA professional lab confirmed it in mammalian cells.
- Preclinical evidenceAnimal and safety studies support it.
- Human clinical evidenceClinical trials in people show it.
How the public metrics are defined
- Research tasks completed
- Research tasks that succeeded during the current UTC day. Platform housekeeping (self-checks, benchmarks) is excluded.
- Papers analyzed
- Papers whose findings were extracted by the Paper Analysis Agent. Every extracted finding must quote the abstract verbatim; code rejects any quote that is not found.
- Papers screened
- Papers whose identity and retraction status were verified against independent bibliographic sources.
- Simulations completed
- Model experiments and model runs that finished (each is reproducible from a recorded manifest).
- Hypotheses evaluated
- Belief updates: a research question's competing explanations re-weighed against the current independent evidence.
- Hypotheses rejected
- Hypotheses whose belief fell below the rejection threshold with enough independent evidence; each is recorded in Failure Memory.
- Compute hours
- Wall-clock time of succeeded tasks, from start to finish, summed.
- AI model calls / cost
- Metered calls to AI models and their cost, from the engine's cost ledger.
- Active agents
- Agents with a task running at the moment of publication.
- Findings under review
- Candidates that cleared the evidence bar and are waiting for a human reviewer.
Research velocity
Weighted count of research operations that succeeded in the last 24 hours.
A higher number means more research work was completed. It does not mean better science, and it says nothing about medical outcomes.