Inside-Out — Invariant-Core Guided Object Formation Through Spectral Affinity and Least-Expense Interior Paths

Inside-Out proposes a core-first approach to visual object formation in which object-relative structure is inferred outward from stable interior references rather than primarily inward from detected boundaries. Its central proposition is concise: Invariant before Boundary. The framework investigates whether a sufficiently stable latent reference — an Invariant Reference Core — can be recovered from visual information and used as the originating reference from which surrounding image structure is evaluated, allowing object identity and membership to emerge rather than be assumed. The proposed computational sequence is: Image → spectral state → candidate cores → anchored affinity → least-expense interior connectivity → core cooperation/competition → object membership → emergent boundary. Version 1 develops the conceptual and mathematical foundations of this approach and reports exploratory experiments using the Arc de Triomphe as a running structural case. The experiments examine spectral core candidates, fixed-core affinity fields, and core-anchored least-expense interior paths. An A/B/C ablation compares local neighbour propagation, direct core affinity, and direct core affinity combined with core-anchored least-expense connectivity. In the reported exploratory tests, the anchored formulation produces less diffuse influence fields across all four tested reference cores. The work also reports generalisation stress tests on ordinary multi-object, textured, and partially occluded scenes. These tests expose an important limitation: a spectrally strong candidate core is not necessarily a semantically correct object core. This negative result motivates explicit mechanisms for core cooperation, competition, rejection, uncertainty, and re-anchoring. Two hypotheses remain deliberately open. First, whether core-conditioned selective computation can achieve comparable visual inference quality with lower computational cost than conventional dense approaches. Second, whether core-first organisation can provide a useful architectural principle for trainable neural systems. Version 1 does not claim that either hypothesis has yet been established. The release includes the Version 1 white paper, experimental figures and datasets, Python code used for exploratory analysis and figure generation, source test images suitable for redistribution, citation metadata, and integrity checksums. Inside-Out extends a research lineage developed through the Invariant Pillar, Invariant Pillar Technical Report, and Adaptive Invariant Pillar (AIP) frameworks while shifting the investigation toward a falsifiable core-first model of visual inference. Author: Mukund K. V.ORCID: 0009-0007-2335-7992Version: 1.0DOI: 10.5281/zenodo.23881266