Findings from running systems. We publish what we observe, not what we speculate.
ALPS evolves Lisp-like programs to predict binary sequences using compression as fitness. Across 88 experiments, ablations showed that crossover, tournament selection, and a module system each had zero measurable impact. The engine's effectiveness comes from two sources: Lamarckian MLE parameter refit and operator injection. Everything else was removed.
The thesis studies a supply-chain threat model built on subliminal learning results from 2024 to 2026. Core claim: deterministic controls protect artifacts, while behavior in foundation models is distribution-shaped and can drift through channels that pass content-level screening.
BlackAgent starts constrained and earns broader access through a trust score that rises with successful task completion and drops with failures. Combined with contract-governed execution and a deterministic state machine for safety decisions, this architecture produced agents that are measurably more persistent, more efficient, and more predictable.
Nadar is a real-time voice companion for visually impaired users that speaks Moroccan Darija natively. Building voice-first AI in a language with no standardized orthography, limited NLP support, and heavy code-switching between Arabic, French, and Spanish required rethinking assumptions that most assistive technology frameworks take for granted.
Across two very different systems, a motion visualization engine and an autonomous AI agent, we observed the same pattern: the data structure chosen to store state had more influence on the system's character than the algorithms that processed it. TRAJEKT's ring buffer depth controls visual personality. BlackAgent's memory layers shape agent behavior. The container matters more than the contents.