The Constellation Calculus:

This project develops and empirically stress-tests the Constellation Calculus, a graphalgebra in which knowledge is represented as "glyphs": graphs whose edges carry complexweights w·e^(iθ) with gauge-theoretic phases, so that the only frame-invariant — andtherefore meaningful — quantities are loop holonomies. Four operators act on thisstructure: gauge precession, measurement with entropy-drawn back-action, node elimination(occultation), and dissipative inter-glyph resonance. The project follows a falsification-native, preregistration-style protocol throughout.The calculus itself is specified as twelve executable claims, each written as a testdesigned to break it; every experiment registers its quantitative prediction in thesource code before the experiment runs; anchor conditions must reproduce prior resultsdigit-for-digit before new conditions are interpreted; and refutations are preserved inthe record rather than revised away. Of five registered experimental predictions, fourwere refuted, and the sequence of refutations is reported as the primary result. Principal findings: (1) a projection theorem for inter-glyph resonance, yielding"interfacial consensus without semantic consensus" — two glyphs can negotiate to perfectphase agreement at their interface while their loop meanings remain permanently distinct;(2) a measured separation between preserving a glyph's meaning under node elimination andpreserving its identity under progressive corruption, traced through a four-stage refinedprinciple (mass fidelity → residue → displacement → margin) to a diversity-preservationcriterion for identity; and (3) an outward-facing classification experiment showing thatholonomy carries task-relevant loop structure (enclosure) that permutation-invariantpooling of the same edge values cannot recover, with an advantage that widens under noise. The project contains the complete reference implementation (Python, NumPy only), theexecutable falsification suite (12/12 passing), all experiment scripts with theirregistered predictions in the module docstrings, the living specification with every rawresults table, and the full manuscript source. All random seeds are fixed; every numberin the manuscript is reproducible from the released code.