MERRCURR: Autonomous Cognitive Self-Modification in Frozen Sub-Frontier Language Models Through Prediction-Verified Reasoning Failure Diagnosis, Quality-Gated Rule Promotion, and Epistemic Gap Identification on Consumer Hardware

This paper presents MERRCURR (Memory-Enhanced Recursive Reasoning with Correction-Updated Reliability Recalibration), an integrated method for autonomous cognitive self-modification in frozen sub-frontier language models deployed on consumer hardware, without fine-tuning, reinforcement learning, weight modification, cloud dependency, or external model involvement. Six stages are demonstrated individually with real-world operational data on a 14-billion parameter frozen model (Qwen3 14B) running on an Apple Mac Mini M4 with 24GB unified memory: prediction-verified self-recalibration, autonomous post-mortem generation, epistemic gap identification, candidate cognitive rule drafting, adversarial quality-gated validation, and meta-epistemic recursive connection. The full recursive loop has not yet completed one end-to-end cycle and remains future work. Third paper in a series on sovereign AI systems on consumer hardware. Prior work: Ground Truth Engineering (Jaber, 2026a) and Calibrated Self-Assessment (Jaber, 2026b).