ALETHEIA is a falsification engine: it models human knowledge as a graph of claims, finds the load-bearing assumptions resting on thin evidence, and computes the single experiment that would remove the most uncertainty from science.
We generate hypotheses faster than ever. We have no infrastructure for refuting them at scale. Popper's falsification, made computational and planetary.
Citation graphs count popularity. They cannot tell you which result is load-bearing — the one study that, if it falls, collapses an entire subtree of accepted conclusions.
Language models summarize what is widely said. That makes fragile, repeatedly-cited beliefs more entrenched, not less. They optimize agreement, not truth.
H. pylori. Continental drift. The literature routinely disagrees with itself for years before anyone notices. No system watches the contradiction fronts.
Nobody can answer: what is the most valuable experiment humanity is not running? ALETHEIA makes that a computable quantity.
A slice of "gut microbiome → behavior" research — a field with documented replication problems. Click any claim to inspect its evidence and fragility. Then ask the engine what to test next.
Each candidate is scored by the expected uncertainty (in nats) it would remove from the entire graph — both possible outcomes, weighted by their predicted probability.
Every claim carries a posterior in log-odds. Evidence shifts it by its log-likelihood ratio × quality. Dependency edges let a doubtful parent erode a child's grounding — belief flows through the structure.
For each piece of evidence, retract it and re-infer. The total movement across the graph is its fragility. High fragility = a planetary belief resting on one shaky study.
For each untested experiment, simulate both outcomes, weight by predicted probability, and measure the expected drop in graph entropy. The top result is the cheapest path to the most truth.
New evidence → recompute confidence → re-rank fragility → propose next experiment → repeat. Knowledge stops being a static archive and becomes a living, self-auditing system.
Bayesian engine in Python + interactive browser demo. Runs real inference on real science.
View Code →Full blueprint: problem, insight, technical foundations, architecture, risks, and implications.
Read Paper →8-slide presentation with editorial identity, engine figures, and proof of concept.
View Deck →"A theory which is not refutable by any conceivable event is non-scientific." ALETHEIA gives that principle an engine — and points it at everything we think we know.