Kronumos 2 Kairos

Cost-bounded automated program repair with a Dual-Brain Cybernetic Sub-Cortex.

Kronumos 2 Kairos pairs a fine-tuned 7B open-weight code model (Qwen2.5-Coder-7B-Instruct, NF4) with the Tokenectomy Dual-Brain Sub-Cortex, a zero-allocation deterministic Rust runtime. The neural layer handles fault localization and patch synthesis. The deterministic layer handles context pruning, secret redaction, POSIX hunk re-anchoring, AST validation, and healing. Patches that fail validation are withheld (Zero Dirty Diff).

Paper: Kronumos 2 Kairos: Cost-Bounded Automated Program Repair via Dual-Brain Cybernetic Sub-Cortex on SWE-bench Verified (Zenodo, v2.0.0)

How it works

1. Prune & sanitize: Issue De-Noiser, AST frame pruning, and linear-time secret redaction (JWT, Bearer, AWS, DB URI, private keys).
2. Synthesize: the 7B model emits a replacement patch from a scrubbed prompt (~2.5K tokens on average).
3. Validate: path grounding, exact-offset matching, POSIX header reconstruction, AST syntax check, AST Auto-Bracket & Indentation Healer, and a git apply dry run.
4. Gate: valid patches are submitted; everything else is withheld as an empty patch.

The loop allows up to 5 self-healing turns. Rejection reasons from the Sub-Cortex are fed back to the model. No test execution feedback is used.

Results: SWE-bench Verified (N = 500)

Official Docker harness (swebench.harness.run_evaluation), 100% container coverage, 0 infrastructure failures.

Ablation: AST Auto-Bracket & Indentation Healer

The healer purged synthetic comments, normalized indentation, and cleaned trailing whitespace. This is a single run on a small sample.

Per-repository results

Evaluation scope and limitations

• File-localized setting. 68.4% (³⁴²⁄₅₀₀) of issues contain a stack trace used to find target files; the rest are localized by lexical references and AST symbol lookup. This measures patch synthesis, not unconstrained repository search.
• No runtime feedback. The model never runs pytest or any harness during inference.
• Low absolute resolve rate. The contribution is cost, token efficiency, and safety, not competitiveness with frontier interactive agents.
• Training data. 1,080 trajectories from PRs outside SWE-bench, filtered by instance ID, commit SHA, and patch hash. Indirect pre-training exposure of the base model cannot be ruled out.
• Per-turn latency and average turn count are not reported for this release.

Quickstart

ollama run hf.co/NadevA23/Kronumos-Kairos-v2-GGUF

import ollama

response = ollama.chat(
    model='hf.co/NadevA23/Kronumos-Kairos-v2-GGUF',
    messages=[{
        'role': 'user',
        'content': 'Analyze this runtime error and generate a unified diff patch:\n'
                   'ZeroDivisionError in calculate_metrics()',
    }],
)
print(response['message']['content'])

Ollama exposes an OpenAI-compatible endpoint at http://localhost:11434/v1 for clients that accept a custom base URL.

Links

• Paper: https://doi.org/10.5281/zenodo.23013104
• Code: https://github.com/Tokenectomy-Labs/Kronomus
• Portal: https://kronumos.eu.org
• Contact: kronumos@proton.me

🤝 Acknowledgements & Quantization Credits

• Original model weights & architecture: Muhammad Naufal Daffa (NadevA23) / Kronumos AI
• High-efficiency GGUF imatrix quantization: Special thanks to Michael Radermacher (mradermacher) for the high-quality importance-matrix quantizations that power this release on edge devices.
• Model weights on Hugging Face: NadevA23/Kronumos-2-Kairos

Citation

@misc{daffa2026kronumos2,
  author    = {Muhammad Naufal Daffa},
  title     = {Kronumos 2 Kairos: Cost-Bounded Automated Program Repair via Dual-Brain Cybernetic Sub-Cortex on SWE-bench Verified},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23013104},
  url       = {https://doi.org/10.5281/zenodo.23013104}
}