lin-0275

For homotopical extensions, Richter [ 2020 ] provides a route from ordinary categories to homotopy theory, while May [ 1992 ] develops the simplicial foundations. Accessible and locally presentable categories [ Adámek and Rosický , 1994 , Makkai and Paré , 1989 ] are natural settings for studying categories of models, completion, and coalgebraic semantics beyond finite sketches. Universal coalgebra [ Rutten , 2000 ] gives a complementary language for unfolding, behavioral equivalence, and coinductive proof.

Higher tangent structure can be approached through Weil-algebra classifications [ Leung , 2017a , 2017c ] ; MacAdam’s functorial treatment connects Weil semantics, Lie theory, nerves, and realization [ MacAdam , 2022 ] . Reverse derivative and reverse tangent categories [ Cockett et al. , 2020 , Cruttwell and Lemay , 2024 ] frame the reverse-transport problem. On the learning side, categorical deep learning [ Gavranović et al. , 2024 ] and copresheaf topological networks [ Hajij et al. , 2025 ] suggest how richer declarations might compile into architectures. The open problem for LINCS is to join these theories to uncertainty-aware, mechanized admission without erasing the domain semantics that make an obstruction meaningful.

For the compression view of learning, Rissanen [ 1978 ] gives the foundational shortest-description formulation and Grünwald [ 2007 ] develops modern MDL in depth. Li and Vitányi [ 2019 ] provides the algorithmic-information background, including the invariance and noncomputability qualifications that separate ideal complexity from an operational code. Zahavy [ 2026 ] challenges compression as an account of scientific invention by separating inductive compression and deductive verification from abductive changes of axioms. The LINCS question is how such codes should descend through declared quotients and compose across covers, and how candidate declaration changes should be generated and audited, without replacing typed admissibility by a single complexity ranking.

McCarthy [ 2007 ] frames a complementary limitation as the gap between appearance and the reality that produces it. His examples connect ordinary object permanence, inverse problems, and the invention of theoretical entities such as atoms. Read together with Pearl’s causal critique, this makes explicit why latent explanation and interventional validation are distinct obligations.

Lenat’s account of AM [ Lenat , 1976 ] remains a foundational study of heuristic concept formation and conjecture generation. The methodological analysis by Ritchie and Hanna [ 1984 ] is a useful counterweight when assessing how much novelty arose from general discovery principles and how much from hand-engineered representation and heuristics. Together they make AM a particularly instructive precursor for LINCS: a system can be strong at proposing extensions while remaining weak at proof, admission, and transfer beyond its initial declaration.

The original papers make the two scientific-discovery loops especially concrete. Planck’s radiation-law paper [ Planck , 1901 ] begins from the empirical failure of a previously derived distribution and changes a constitutive assumption in response. Einstein’s systematic presentation of general relativity [ Einstein , 1916 ] derives light deflection as a consequence of the new frame, while the eclipse report of Dyson et al. [ 1920 ] records the independent observational test. They are useful not as myths of solitary inspiration, but as paired examples of obstruction-driven re-declaration and prediction-driven experiment design.

Malthus’s population essay [ Malthus , 1798 ] , Darwin’s later recollection of reading it [ Darwin , 1958 ] , and the mature argument in On the Origin of Species [ Darwin , 1859 ] document a different discovery pattern: transport of a mechanism across domains. The joint Darwin–Wallace communication [ Darwin and Wallace , 1858 ] is essential historical context and guards against turning a distributed scientific development into a story of isolated inspiration.