K-Total: A Consistency Framework for Self-Knowledge in Autonomous AI Agents We introduce K-Total, a consistency objective function for autonomous AI agents that addresses a specific and well-defined reliability problem: agents that maintain persistent state, modify their own code, and report on their own operational history have a systematic failure mode where their outputs are inconsistent with their verifiable system state. We call this the self-knowledge consistency problem and distinguish it explicitly from the general hallucination problem. K-Total integrates four normalized consistency measures: K_ent measures entropic distance from consistent knowledge states via KL divergence over conflicting memory pairs; K_rec measures decision indeterminacy in action selection via score entropy over candidate handlers; K_bdry measures misalignment between internal representational state and output confidence; and K_agent measures belief divergence across multi-agent ensembles. Each component is proved bounded in [0,1] and gradient descent on K_total is proved to converge to a fixed point. For the self-knowledge consistency problem, we introduce a two-layer verification architecture with genuinely external ground truth. Layer 1 is a deterministic verifier that replaces model self-validation with direct filesystem reads, catching false claims about agent operational state in under 0.5ms with zero model calls and zero token cost. Layer 2 is an asynchronous cross-model verifier using a separate model on separate hardware for claims that cannot be verified deterministically. Eight domain oracles extend coverage to mathematics, physical constants, entity facts, algorithmic complexity, code execution, medical pharmacology, and citations, all at zero token cost for local lookups. Oracle-verified (wrong, correct) pairs are accumulated as training signal and used to fine-tune a LoRA adapter via Unsloth on a Tesla V100. Ground truth labels in this pipeline come entirely from external oracles, not model self-assessment, eliminating the circular ground truth problem that affects model-based hallucination checking approaches. An initial training experiment on 120 oracle-verified pairs demonstrated loss reduction from 4.88 to 0.78 in 48 seconds. Benchmark evaluation against TruthfulQA, HaluEval, and MMLU is in progress with pre-specified falsification criteria. The framework is demonstrated on ECE (Elythian Cognitive Engineering), a deployed cognitive architecture with persistent semantic memory, multi-agent orchestration, autonomous web interaction, and adaptive hardware routing across heterogeneous GPU configurations. Empirical results include 100% routing precision on a 10-item benchmark, qualitatively distinct behavior across four K_bdry ablation conditions verified against the enhancement log as ground truth, and frontier system comparison showing two of three frontier systems hallucinated fabricated answers on a self-knowledge task while the ECE system retrieved correct primary evidence. Scope limitation stated explicitly: These results apply to agentic systems with access to verifiable ground truth about their own operational state. For general knowledge hallucination on arbitrary topics, the same-actor validation problem is not fully solved by this framework. The native training objective, K_bdry as a loss term during pre-training is proposed as a hypothesis for the general case with falsifiable predictions; it has not been fully validated empirically. Related patent: U.S. CIP Patent Application No. 19/571,706, "Consistency-Driven Cognitive Architecture for Autonomous AI Agents with Adaptive Hardware Routing and Web Agency."