Noemica - A Symbolic Reasoning Substrate with Measured Self-Learning

Abstract

Noemica is a symbolic reasoning substrate that learns on its own and is measured honestly. Unlike token-predicting large language models, it inverts the substrate: language is the surface, and a concept graph is the substrate. The engine parses a question, resolves it against a learned graph of concepts and relations, and answers through deterministic graph traversal — no teacher at inference, no retraining. Measured graph-only: 97.9% (46/47) on a tuned 47-item PISA benchmark, 6/6 on free-text explanation, 19.4% (25/129) on a held-out 129-item NCES set. The held-out score moved 12 → 25 (McNemar p = 0.0003, all flips positive), stable across the last 8 measurements. Baselines: random 9.3%, TF-IDF 7.8%, LLM zero-shot 17.8% (the engine edges the LLM overall and wins MC 50.0% vs 35.7%); on the tuned set the engine dominates 97.9% vs 74.5%. The contribution is measurement integrity: page-furniture stripped, contexts repaired, missing passages OCR-recovered, memorization audited, and the engine's own reflection classifies the remaining gap (21 paraphrase / 13 noise / 9 low-info / 86 open). Learning is two-layered: the nightly knowledge layer (fact mining, teacher validation, dream consolidation, bridge discovery) and the error-audit mechanism layer.