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

zenodo.org · 2026-10-02 · by Long Nguyen ·generic
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.

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

Published September 28, 2026 | Version 2.0.0

Preprint Open

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

Authors/Creators

Description

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.

Files

###

Kronumos\_Paper\_v2.pdf

###

Files (301.9 kB)

| Name | Size | Download all | | ---------------------------------------------------------------------------------------------------------------------------------------------- | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Kronumos\_Paper\_v2.pdf md5:b027d42db427754e559bcab4d2a6e38c | 301.9 kB | Preview Download |

Jump up

zenodo.org · 7 images · 40 links

Comments (0)

No comments yet