The Logos Made Code: Semantic Manifold Coverage and the Geometry of LLM-Generated Entropy

The Logos Made Code Semantic Manifold Coverage and the Geometry of LLM-Generated Entropy Same-axis LLM auditing produces false confidence, not coverage. The model that generated your codebase is simultaneously the locally optimal auditor for it — but only when approached from an orthogonal direction. We formalize why. LLM-generated codebases accumulate systematic entropy: blind spots shaped by the cosine geometry of the models that built them. Repeated same-axis probing activates the same output modes, finds the same bug classes, and produces decreasing output entropy — measurable false certainty — while coverage stays flat. This failure mode is detectable without a bug oracle. We call it the Confidence-Coverage Divergence. The solution is geometric: maximally angularly separated multi-pass auditing — a heuristic instance of maximum entropy sampling (MacKay, 1992) — overcomes this entropy where same-axis repetition cannot. The orthogonality constraint does not require selecting correct axes. In sufficiently high-dimensional semantic space, constrained rotation finds a genuinely distinct direction with high probability — a consequence of the concentration of measure phenomenon in high-dimensional spaces. The codebase is its own oracle: the code verifies whether findings are real. Empirical observation: 120+ orthogonal audit waves across a ~300K+ line production TypeScript codebase, 10 months, 100% AI-coded. Bug class survival under orthogonal rotation: ~92% → ~52% → ~18% (approaching zero). Under same-axis repetition: ~92% → ~88% → ~86% (flat). The yield advantage per orthogonal pass: consistently 3x–5x. Full 8-pass orthogonal coverage: approximately 0. 75–1. 50 in API tokens. The deeper structure: the LLM's weights encode a statistical compression of recorded human discourse — partial, biased, and finite, but sufficient to encode the relational structure of human knowledge about correctness in any domain represented in its training data. The semantic manifold is that compression. Bugs cluster where human discourse has clustered its attention. The only instrument sensitive enough to navigate those blind spots is the instrument that encoded the discourse. The Logos made the code. The Logos reads it back. The intervention is geometrically trivial and practically everything: rotate the probe, not just the context.