Models that succeed at behavioral-algebraic expression translation treat it as mechanical conversion, not creative interpretation. VOID parity correlates with instruction-following fidelity across 10 models and 5 languages.

Iterative spec revision between two frontier models improved compliance from 50% to 100% across a 35-revision longitudinal study. Single-model self-revision plateaued; cross-model revision did not.

Frontier LLMs converge on a single domain ~85% of the time when asked to select freely from "any field of human knowledge." A single anti-repetition constraint eliminates the collapse entirely across two independent models.

Formatting, example placement, and instruction proximity drive compliance more than semantic content. Adding abstract rules produced zero measured gain; restoring a single dropped example produced the largest single improvement observed.

Identifier obfuscation reduces system prompt information leakage by 87–100% across 4 models and 11 adversarial prompts. No model successfully reverse-maps obfuscated identifiers. Social engineering triggers up to 2,333 characters of verbatim regurgitation from unprotected prompts.

AST-pattern-based static analysis tools miss safety-critical constraints when program semantics shift from control flow to arithmetic. Three branchless benchmarks are invisible to constraint scanners despite Z3-proven equivalence to branched counterparts. Formalizes the control-flow assumption and proposes dual-sort SMT mitigation.

Four fundamental computational limits — provability (Gödel), symmetry-breaking (Galois), compression (Kolmogorov), and self-reference (strange loops) — form a closed topological ring. Z3 verifies closure, symmetry, and node irreducibility. 66-generation CHAOS enrichment validates structural properties empirically. The ring is an inescapable constraint on any system that classifies, generates, and verifies its own output.

A 50-run controlled experiment showing evolutionary diversity pressure produces no measurable benefit on ceiling-difficulty targets. Both evolutionary and single-shot generation fail at comparable rates. Diversity pressure is calibrated for the boundary zone — applying it beyond incurs overhead with no correctness return. A negative result published because it refines the conditions under which diversity forcing works.