Deterministic governors for real-time AI.

Lioren is building AEC (AI Entropy Control), a deterministic governor for language models. AEC detects the conditions that produce hallucinations during inference, and intervenes to prevent the bad response from reaching the user.

AEC is deterministic: the same input and governor state produce the same response, making its decisions reproducible and auditable rather than probabilistic. It runs in-stream during generation, not as a post-hoc filter on completed output, so the hallucination is caught before it is emitted.

AEC is not a wrapper, prompt-engineering layer, or retrieval system. It is a separate governance process that reads from the model and intervenes before hallucinated content reaches the user.

Designed for self-hosted language model deployments.

We build governors against measured model behavior, not assumed model behavior.

AEC is being designed in response to empirical probing of how language models actually fail. We construct controlled prompts, run them across open-weight models with multiple iterations per prompt, and capture quantitative measurements of where the models hallucinate and how.

The current probe focuses on hallucination patterns in completion-style prompts, using historical art criticism as source material. These findings inform AEC's design.