⚡ Kronumos 2 Kairos: The Dual-Brain Sub-Cortex Autonomous Program Repair Engine

Organization: Tokenectomy Labs
Base Model: Qwen/Qwen2.5-Coder-7B-Instruct
Research Preprint: Springer Nature Research Square (DOI: 10.21203/rs.3.rs-11205335/v1)
Lead Author: Muhammad Naufal Daffa (ORCID: 0009-0000-7909-4916)
GitHub Repository: Tokenectomy-Labs/Kronomus

Kronumos 2 Kairos is an open-weight 7B cybernetic autonomous program repair (APR) model fine-tuned for high-precision code remediation on real-world production software bugs. It pairs parametric neural intuition with a deterministic, zero-allocation Rust Sub-Cortex (libtokenectomy_subcortex.so, C-ABI 5µs latency).

🏛️ The Kronumos Dual-Brain Family Portfolio

🥊 Benchmark Verification: SWE-bench Verified (500 Instances)

Evaluated end-to-end on the official Princeton SWE-bench Verified benchmark (500 production instances across Django, Scikit-Learn, PyData Xarray, Sphinx, Sympy, etc.) using official Docker execution containers.

🏆 Verified Resolved Production Issues:

1. django__django-13569: Broken aggregation expression logic in database queries.
2. django__django-13658: Management command argument parser collision.
3. django__django-14855: Admin URL generation prefix regression.
4. django__django-15104: Model custom key migration constraint hazard.
5. django__django-16333: Many-to-many relationship foreign key mapping.
6. pydata__xarray-4629: Multi-index coordinate slice dimension regression.
7. scikit-learn__scikit-learn-10844: Pipeline estimators parameter validation fault.
8. sphinx-doc__sphinx-8595: Python domain autodoc signature formatting error.

🔬 The Dual-Brain Cybernetic APR Architecture

Traditional LLM agents rely exclusively on multi-turn prompt loops, generating massive token overhead and hallucinating syntax formatting. Kronumos Kairos decouples cognition into two integrated computing cortices:

           [Raw GitHub Issue Discussion]
                         │
                         ▼
        ┌──────────────────────────────────┐
        │   Sub-Cortex IssueDeNoiser       │
        │   - Excises human chatter/quotes │
        │   - Tags user reproduction code  │
        │   - Extracts Core Signal Triad   │
        └────────────────┬─────────────────┘
                         │
        [Cleaned Technical Specification]
                         │
                         ▼
        ┌──────────────────────────────────┐
        │  Procedural Cognitive Kernel     │ ◄─── Procedural Seeds (<64 bytes)
        │  - BoundaryCondition Invariants  │      (Zero-DB L1 Cache Execution)
        │  - DefensiveNullWrap / PopGuards │
        └────────────────┬─────────────────┘
                         │
    ┌────────────────────┴────────────────────┐
    ▼                                         ▼
┌───────────────────────┐         ┌───────────────────────┐
│   CORTEX (Neural)     │         │ SUBCORTEX (Deterministic)
│  Qwen2.5-Coder-7B     │ ◄─────► │ Tree-sitter AST Slicer│
│  - 5-Step CoT Reason  │ Dual-Key│ Auto-Bracket & Indent │
│  - Precise Code Hunk  │ Consens.│ Merkle Causal Ledger  │
└───────────────────────┘         └───────────────────────┘

1. Issue De-Noiser: Strips human conversational chaff, extracting the core reproduction triad.
2. Procedural Cognitive Kernel: Diagnoses invariants across 9 domains (Boundary, Defensive, Concurrency, etc.) without external database lookups.
3. Dual-Key Consensus Gate: Requires simultaneous semantic approval and deterministic AST validation before admitting state changes.
4. Auto-Bracket & Indentation Healer: Deterministically balances parentheses and enforces strict PEP 8 4-space block indentation.

⚡ Native Rust Sub-Cortex Runtime

This repository includes the precompiled native Linux x86_64 binary libtokenectomy_subcortex.so and Python C-ABI bridge tokenectomy_subcortex_rust.py.

from tokenectomy_subcortex_rust import RustSubCortex

subcortex = RustSubCortex()

# 1. Clean noisy issue descriptions (93.5% token reduction)
clean_spec = subcortex.denoise_issue(raw_github_issue)

# 2. Heal indentation drift and unbalanced brackets with sub-microsecond latency (5µs)
healed_code = subcortex.heal_indentation(candidate_code, base_indent=4)

💻 Quickstart: Running Inference

With Transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "NadevA23/Kronumos-Kairos-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are Kronumos Kairos, an expert autonomous program repair engine."},
    {"role": "user", "content": "Fix the issue in the following function:

def safe_divide(a, b):
    return a / b"}
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Quantized GGUF (Llama.cpp / Ollama):

For quantized local execution on consumer hardware, visit NadevA23/Kronumos-Kairos-v2-GGUF.

📜 Citation

@article{daffa2026kronumos2,
  author    = {Muhammad Naufal Daffa},
  title     = {Kronumos 2 Kairos: Cost-Bounded Automated Program Repair via Dual-Brain Cybernetic Sub-Cortex on SWE-bench Verified},
  journal   = {Research Square},
  year      = {2026},
  doi       = {10.21203/rs.3.rs-11205335/v1},
  url       = {https://doi.org/10.21203/rs.3.rs-11205335/v1}
}

License: Apache 2.0
Maintained by: Tokenectomy Labs