The Dagger in the Garden (Revised Edition) A Narrative Scientific Report on Complex-Valued Attention, Phase Interference, and the Learning of Ambiguity in a Single Poem

Abstract

This paper reports a complete end-to-end experiment in complex-valued language modeling. We take the mathematical theory developed in The Dagger and the Query: A Unified Mathematical Theory of Complex, Unitary, and Quantum Attention (Suansook, 2026) and instantiate it as a working system: a character-level ComplexGPT trained on a single 595-character poem, The Rose Is Not a Rose. The poem was chosen because it exhibits simultaneous lexical, syntactic, semantic, and pragmatic ambiguity --- precisely the structure that the imaginary part of a complex attention score is designed to represent. We make three contributions. First, we prove the mathematical and implementation correctness of the complex attention machinery: all six verification tests pass, including the interference identity \( O_i = 2w\cos\theta \cdot v\ \) for symmetric phases, and the closed-form parameter count \( P_{\mathrm{non\text{-}emb}} = 4Vd + L(4d + 24d^2) + 2d \) matches the observed count of 53,824 for \( V=34,\ d=32,\ L=2 \). Second, we demonstrate trainability: training loss falls from 3.55 to 1.44, and validation loss falls from 3.56 to a minimum of 2.54 before rising mildly --- a signature of overfitting on a small corpus, not of architectural failure. Third, we produce recognizable English with the poem's vocabulary (rose, wind, garden, word, the, and), though not cohesive poetry. We are explicit about what this paper does not show. It does not prove that complex attention outperforms real-valued attention. No real-valued control baseline was trained. It does not measure whether the imaginary part of the attention score is actually used — no diagnostic was instrumented, so we cannot say whether the model has collapsed to real-valued attention or is genuinely exploiting phase. It does not establish that phase carries linguistic meaning. These are empirical questions that remain open. We position the paper as a software and mathematical proof-of-concept, not a machine learning hypothesis paper. Its contribution is correctness and trainability, not benchmark superiority. We close by specifying the experiments that would elevate it to the latter: (1) a real-valued control baseline, (2) an instrumented phase diagnostic, (3) a larger and more diverse corpus, and (4) a comparison of validation loss and sample quality across the two architectures. Until those are run, the paper's central hypothesis — that phase-preserving attention provides an advantage on phase-sensitive tasks — remains unproven.