Navigational Intelligence: An Information-Theoretic Limit on What AI Can Learn from Published Research

Every contemporary AI system learns from the published outcomes of research — papers, preprints, finished text — and therefore learns what researchers concluded, not how they got there. We prove that this omission is a limit, not an engineering gap. Under a single explicit assumption — that the publication map taking research trajectories γ to outcome documents E_T is non-degenerate (distinct trajectories can produce the same outcome with positive probability) — Shannon's data processing inequality yields I(γ) > I(E_T): the conditional information ΔI = I(γ | E_T), which we call the navigational intelligence, is strictly positive and irreducible by any algorithm operating on outcomes alone, at any scale, with any architecture. We identify three observable categories of this gap — dead ends, corrections, and breakthroughs — and prove an exact chain-rule decomposition of ΔI into these categories plus a completeness term, with non-negative conditional-dependence residuals. We further prove an information hierarchy over a proposed preservation infrastructure, the Living Research Graph: papers, trajectories, connected trajectory pairs, and the full graph each strictly exceed the level below. The framework's central open problem is geometric: construct a metric on epistemic states under which the counterfactual reference trajectory is a geodesic. We state three candidate constructions (Fisher–Rao, embedding pullback, Nielsen-type), two correctness conditions any candidate must satisfy, and a falsifiable empirical test. The limit is not a data-quantity problem but a memory-architecture problem: current AI encodes declarative knowledge (what was found) and discards evolutionary knowledge (how the finding was reached). AI should remember how knowledge was created, not only what was created.