Kronumos 2 Kairos: Cost-Bounded Automated Program Repair via Dual-Brain Cybernetic Sub-Cortex on SWE-bench Verified

Automated Program Repair (APR) in authentic software repositories presents a formidable challenge: models must localize defects within repository contexts, synthesize character-exact syntactic modifications, and preserve regression invariance without corrupting preexisting functionality. In this paper, we present Kronumos 2 Kairos, a cost-bounded autonomous program repair engine coupling a fine-tuned 7B open-weight code model (Qwen2.5-Coder-7B-Instruct) with the Tokenectomy Dual-Brain Cybernetic Sub-Cortex—a zero-allocation deterministic runtime featuring an Issue De-Noiser, a 9-domain Procedural Cognitive Kernel, an Interlocking Causal Invariant Mesh (ICIM), a Dual-Key Consensus Gate, a Cryptographic SHA-256 Merkle Causal Chain, an Immune Cluster Pattern Classifier, and an AST Auto-Bracket & Indentation Healer. Evaluated against the complete 500-instance Princeton SWE-bench Verified dataset using the official Docker testbed with 100% container execution coverage, Kronumos 2 Kairos synthesized 442 candidate patches while withholding 58 (11.6%) via deterministic structural validation (Zero Dirty Diff). The system verified 8 production bug resolutions across four distinct open-source ecosystems (Django, Scikit-Learn, PyData Xarray, and Sphinx) at strictly $0.00 marginal inference cost and an average of 2,512 tokens per task (a 93.5% token reduction vs. industry multi-turn baselines). Furthermore, an empirical ablation demonstrates that the Sub-Cortex AST Healer alone rescued 2 additional resolutions (+33% improvement over the pre-healing baseline), confirming the architectural necessity of deterministic post-processing in open-weight APR pipelines.