CAEL: A Non-Neural CPU Assistant Built from Symbolic, Count and Cellular-Automaton Engines, with a Scoped Comparison against Sub-1B Language Models

We describe CAEL (Cellular Automaton Event Language), a CPU-only assistant that contains no neural network. It combines a cellular-automaton story model, bounded integer session memory, packet-propagation automata for world-state and belief tracking, a Horn-rule graph automaton, count-based language components, retrieval, exact arithmetic, and a bounded Python AST interpreter that checks synthesised programs. On CAEL-CMP-R2-20261028, a frozen set of 80 locally authored synthetic items (memory, conversation, Python, controls), CAEL 0.3.7 scored 57/80 versus 43/80 for Qwen3-0.6B and 41/80 for Qwen3.5-0.8B, and 15/18 on functional Python versus 6/18 and 3/18. Both Qwen models beat CAEL on the conversation category, and on an older, broader 400-question test Qwen3.5-0.8B leads CAEL 181 to 121; Qwen is better at conversation. AI coding agents wrote the 80 items, which overlap CAEL's development families, and also reviewed the answers (unblinded); there was no independent human adjudication, so the result is a scoped confirmation, not evidence of general superiority. CAEL is source-available, non-commercial (CAEL Non-Commercial License). This manuscript was prepared with the help of an AI assistant (Grok Bot) for the author.