SYGON is a hyperdimensional semantic reasoning engine embedded in TauGuard's intelligence layer. Where large language models produce text that sounds correct, SYGON measures whether meaning is coherent, stable, and within authorised semantic boundaries before any output is acted upon.
SYGON builds a φ-lattice — a Fibonacci-spaced embedding space where every concept occupies a mathematically determined position. Tokens are not assigned random vectors. Their positions are derived from SHA-256 hash expansion, modulated by golden ratio geometry, and anchored to domain-specific semantic fields.
When new information enters the system, SYGON does not evaluate it linguistically. It measures it geometrically — against the established structure of the φ-lattice. If coherence falls below threshold, SYGON refuses the pipeline: structurally, not probabilistically.
How tightly incoming tokens cluster within established semantic fields. The geometric closeness of new meaning to existing structure.
Information disorder of the current cognitive state. High entropy signals semantic instability — a precursor to incoherence.
Fraction of semantic capacity under active computation. Load spikes indicate contested meaning or boundary ambiguity.
Proportion of the known semantic space currently activated. Low coverage means reasoning is operating in unexplored territory.
Tokens or patterns that deviate from established field geometry. Non-zero anomaly count triggers escalating verification.
Every token occupies a mathematically determined position. Fibonacci-spaced. Golden ratio geometry. Toroidal topology — no edges, no artificial boundaries.
Five interlocking mathematical structures that make SYGON's verification deterministic rather than probabilistic.
Fibonacci-spaced substrate. Every token placed by golden ratio geometry. The lattice is toroidal — semantic space has no edges, no artificial boundaries. Meaning wraps continuously through the space.
Token positions are not learned. They are derived: SHA-256 hash expansion, modulated by φ, anchored to domain fields. Position is identity.
Each token is assigned a unique prime number as its identity. Compound meaning is computed as prime products. Shared meaning is resolved by greatest common divisor.
Disambiguation by prime factorisation — deterministic, not probabilistic. When two concepts share meaning, the proof is arithmetic, not statistical.
Natural semantic boundaries with no hyperparameters. Every token owns its Voronoi cell — the region of semantic space closer to it than to any other token.
Boundary distance determines ambiguity score. Small boundary distance: ambiguous. Large boundary distance: firmly anchored. Ambiguity is geometry, not guesswork.
Healthy semantic evolution follows a golden spiral trajectory. Concepts grow and shift, but they maintain the φ-ratio relationship between successive positions.
Drift is deviation from that spiral. SYGON distinguishes normal concept evolution — the natural movement of meaning over time — from semantic contamination, which breaks the geometric relationship.
Contextual boundaries that breathe. When reasoning is focused — a precise query about a narrow domain — τ-circles contract, tightening the semantic perimeter to maximise precision and exclude peripheral associations.
When context broadens — a cross-domain analysis, a wide coverage requirement — τ-circles expand to accommodate the necessary field coverage. The breathing rate is derived from τ/φ: the ratio of the circle constant to the golden ratio. SYGON's context awareness is not heuristic. It is rhythmic.
Meaning is not static. SYGON models the coherence field as a wave function — superposition of semantic states, interference patterns, and stable interpretation through coherence resonance.
SYGON is wired into the TauDIL assessment pipeline at three mandatory checkpoints. Incoherent content cannot pass any gate.
Before new entities are committed to the knowledge graph. SYGON measures the coherence delta between existing domain knowledge and incoming content. A drop exceeding threshold triggers rejection.
Contaminated knowledge never enters the authoritative source. The knowledge graph remains geometrically coherent at all times.
Pre-Commit · Knowledge GraphAfter domain rules are loaded, before the AI provider is called. Verifies that the retrieved rule set is semantically coherent and that query tokens have unambiguous dominant context within the φ-lattice.
Incoherent or ambiguous rules cannot reach the AI provider. The AI operates only on verified, geometrically coherent inputs.
Pre-Inference · Input VerificationAfter AI output, before any verdict is committed. Checks output tokens against the authorised semantic context using Voronoi node membership verification. Tokens appearing in output that were never present in the source context are flagged as scope violations.
Critical drift triggers active alignment before the verdict is recorded. The AI cannot introduce meaning that was not authorised by the input context.
Post-Inference · Output VerificationWhen SYGON analyses a token sequence it returns a full cognitive state report — available in real time via /api/sygon/state.
The report is not a classification. It is a geometric description of the current state of meaning — where tokens sit in the φ-lattice, how they cluster, what semantic fields they activate, and whether the output of reasoning has stayed within its authorised boundaries.
Live operational view of the SYGON semantic coherence engine. Real-time wave activity, context ring, Voronoi cell state, drift alerts, and token observation lineage.
SYGON monitors meaning over time — not just at a single instant. Drift is the deviation of a concept's trajectory from its ideal φ-spiral path. SYGON classifies, quantifies, and acts on it before it contaminates downstream reasoning.
Every token in the φ-lattice has a natural trajectory: as context evolves, concepts shift — but healthy semantic evolution follows the golden spiral. Each step rotates by the golden angle and scales by φ. The geometry of healthy meaning is predictable.
Drift occurs when a concept deviates from that trajectory. SYGON computes the angular and radial deviation from the expected φ-spiral position at each evaluation step. When deviation exceeds the τ-circle boundary for that concept's Voronoi cell, SYGON classifies the event as contamination rather than evolution.
Riemannian geodesic navigation with φ-decay step size. Semantic path reasoning that is simultaneously geometrically governed and knowledge-grounded. Prior art does not contain this.
The ManifoldWalker is SYGON's semantic path reasoning algorithm. Given two concepts in the φ-lattice, it does not query a lookup table or traverse a pre-indexed graph. It walks the manifold — navigating the curved surface of semantic space along geodesic paths, with each step governed by φ-decay.
The walker does not move through flat Euclidean space. Semantic space in SYGON is curved — the φ-lattice has Riemannian geometry, where the shortest path between two concepts follows the local curvature of meaning. The walker computes the geodesic: the path of minimum semantic distance on the manifold surface.
Step size at each iteration k is φ−k × the initial geodesic distance. The walk is self-similar at every scale — the geometry of the step is the geometry of the whole path.
At each step, the walker knows which Voronoi cell it occupies. Cell boundaries define semantic ownership — which concept dominates this region of the manifold. As the walker crosses a Voronoi boundary, it records the transition: this is a semantic edge, a relationship between two concepts in the knowledge graph.
The sequence of Voronoi cells traversed during a walk produces a path through semantic space that is both geometrically continuous and conceptually grounded — the cells are meaning, not just geometry.
Each Voronoi boundary crossed during the walk is matched against the Contextual Knowledge Graph. The CKG contains typed relations between concepts — causal, temporal, definitional, evidential, contradicts. When a walk step crosses a boundary that maps to a CKG relation, that relation type is annotated onto the walk path.
The result is a walk that is not just a sequence of positions — it is a sequence of typed semantic moves. The path from concept A to concept B becomes a structured argument: A causes B, B temporally precedes C, C evidentially supports D.
Two walkers are launched simultaneously — one from each concept — converging toward each other to find a semantic meeting point. The meeting point is the position of minimum combined geodesic distance from both sources: the geometric midpoint on the manifold.
Applied to knowledge graph navigation with typed relation annotation, the bidirectional walk produces something that does not exist in prior retrieval or reasoning systems: a geometrically governed path between two knowledge graph nodes where every step carries a typed semantic relation, and the path itself is determined by the curvature of meaning — not by pre-indexed traversal rules.
Clarity of scope is a design principle. SYGON's power comes from doing one thing with mathematical precision.
SYGON produces no output content. It evaluates content produced by others.
SYGON does not have a language model. It has a geometric model of meaning.
Sentiment is irrelevant to coherence. Meaning is verified geometrically, not tonally.
SYGON does not assign labels. It measures geometric relationships in semantic space.
SYGON verifies whether meaning is coherent with what the system already knows, whether the source of that meaning has authority to introduce it, and whether the output of reasoning has stayed within the semantic boundaries of its authorised context.
Every output that acts on the world must first be verified to mean what it claims to mean.