Marr's ghost

David Marr's tri-level analysis has organized cognitive neuroscience for four decades. We argue that it has also constrained it. The framework rests on three assumptions: that cognitive goals can be specified a priori, that neural processes implement algorithms, and that physical substrates are interchangeable realizers of abstract functions. Each assumption was defensible in 1982. Each is now demonstrably false. Slime molds, planarian flatworms, bioelectric networks in non-neural tissue, and degenerate neural circuits pose a dilemma the framework cannot resolve: either extend computational description until it covers everything and therefore explains nothing, or acknowledge that adaptive behavior arises through non-algorithmic physical dynamics. We document how the framework has absorbed anomalies through terminological expansion rather than theoretical revision - a pattern better described as unfalsifiability dressed up as flexibility. Tinbergen's four questions expose what was abandoned: developmental and evolutionary explanation disappeared inside a single privileged level. What remains when that level falls is an obligation rather than chaos: to treat physical dynamics, morphology, and information as co-constitutive of adaptive behavior. The right question concerns how living physical systems - bounded, embodied, evolved - generate the adaptive behavior we have too long called computation for lack of a better word, rather than what algorithm the brain implements.