The Symboken Theory

Tokenization usually treats text as a sequence of orthographic units and leaves phonological structure to a learned model. This paper proposes a different systems boundary. The International Phonetic Alphabet (IPA) is placed in a privileged lower layer (Ring-0), UTF-8/Unicode is preserved as an orthographic transport layer (Ring-1), and a fixed discrete-cosine-transform (DCT) algebra is inserted between them as Ring-$\mu$. Ring-$\mu$ is not a learned model and does not decide semantic truth, policy compliance, or pronunciation correctness. It induces a commutative ring on finite phonological feature vectors by pulling coordinatewise multiplication back through an orthonormal DCT. Nonzero zero-divisor pairs in this ring are defined as \emph{absolute disconnections}: their DCT supports are disjoint. A tangent-chart pole provides a local, directional manifestation of such a disconnection, while the zero-divisor condition supplies the chart-independent invariant. The architecture separates two components that are often conflated. Ring-$\mu$ is fixed, deterministic, and algebraic. A distinct $\mu$-buffer is volatile and stores only temporary alignment and provenance needed to transport spectral marks from IPA positions to UTF-8 byte and grapheme spans. The resulting ``mark-and-screw'' operation marks singular spectral modes and screws those marks through a time-dependent alignment without inserting irreversible boundaries into either representation. The paper provides formal definitions, elementary proofs, a CPU-oriented reference design, falsifiable hypotheses, and an experimental protocol. The proposal is a research architecture, not a validated safety mechanism; its intended contribution is a small, auditable lower layer that can expose phonological discontinuities and token-boundary crossings without requiring an upper language model to understand IPA.