interventional · probabilistic · self-auditing · over 691 claims
A reasoning layer over the unified graph that assumes, propagates, and audits — with the rigorous and modeled parts drawn apart, not blurred
Three capabilities on one probabilistic-implication graph. Intervention — assume a conjecture proved and propagate forward only (Pearl's do-operator): pure deductive reachability, exact. Observation — take a result as evidence and update its causes too: exact by weighted model counting on small cones, approximate beyond — and the approximation is validated against exact, with its error shown. Consistency — audit the graph against itself and triage any claimed new edge before expensive verification: rigorous graph checks. The probability layer is explicitly model-relative: it reports what a coherent Bayesian with these priors and one independence assumption would conclude — never ground truth. Priors are sourced where a basis exists (the P≟NP poll, numerical-verification bounds, documented expert consensus) and honestly tier-labeled estimate where none does. One invariant is enforced throughout: no established result proves an open problem.