Kronumos Architecture & Local Execution
Technical guide for running Kronumos models locally and integrating the deterministic Rust Sub-Cortex kernel for verified software repairs.
Local Inference via Ollama CLI
Kronumos weights are publicly hosted on Ollama Hub for instant, local, and air-gapped program repair. Run the model directly on your workstation without cloud dependencies or external network requests:
ollama run kronumos/Kronumos-2-kairos
Once loaded into Ollama, provide a bug reproduction, failing stack trace, or target file context:
>>> Issue: Prevent unhandled panic when x-signature header is absent
--- a/src/auth/verifier.rs
+++ b/src/auth/verifier.rs
@@ -14,2 +14,3 @@
- let sig = req.headers().get("x-signature").unwrap();
+ let sig = req.headers().get("x-signature").ok_or(AuthError::MissingSignature)?;
Hugging Face Model Weights
Full model checkpoints and quantized GGUF weights are distributed on Hugging Face under the official organization:
Available in 4-bit, 8-bit quantized GGUF format for llama.cpp and Ollama, alongside full precision bfloat16 safetensors for vLLM and TensorRT-LLM clusters.
Hardware Requirements
| Deployment Tier | GPU / Accelerator | RAM / VRAM | Target Workload |
|---|---|---|---|
| Local Workstation (GGUF Q4) | Apple Silicon M-Series / RTX 3060+ | 16 GB Unified / 8 GB VRAM | Local developer workflows, fast AST healing |
| High-Precision (GGUF Q8) | RTX 3080 / RTX 4070+ | 12 GB - 16 GB VRAM | High-precision local program repair |
| Dedicated Node (Bfloat16 FP16) | RTX 3090 / RTX 4090 / A100 | 24 GB+ VRAM | vLLM server, concurrent CI/CD pipelines |
Deterministic Rust Sub-Cortex Kernel
Unlike traditional LLMs that emit unpredictable free-form text, Kronumos pairs neural reasoning models with a deterministic native Rust compiler Sub-Cortex.
Before any patch is accepted, the Rust kernel executes Tree-sitter AST validation in under 5 microseconds via zero-allocation C-ABI foreign function interfaces:
- Zero Syntax Breakage: Indentation errors and missing bracket pairs are deterministically repaired.
- Zero Scope Pollution: Guards prevent destructive function rewrites and preserve docstrings, licenses, and comments.
- Local Secret Redaction: AWS access keys, GitHub personal tokens, JWTs, and database URLs are scrubbed locally before prompt dispatch.
Tree-sitter AST Sentinels
AST Sentinels enforce strict boundary conditions on synthesized diffs. They parse the target code into concrete syntax trees (CST) and guarantee:
AST sentinel checks execute in guaranteed linear time, immune to Regular Expression Denial of Service (ReDoS) hazards.
If AST validation fails or compilation breaks, git working tree state rolls back automatically to zero dirty diff.
POSIX Unified Diff Synthesis
Synthesized repairs are formatted strictly as POSIX unified diffs compliant with standard git tooling. Diffs can be audited and applied using standard terminal commands:
git apply --check fix.patch
git apply fix.patch
Princeton SWE-bench Verified Telemetry
Empirical evaluations are audited against the official Princeton SWE-bench Verified benchmark inside isolated Docker containers:
- Inference Latency: 52.19 seconds average on a single NVIDIA A100-SXM4 40GB accelerator.
- Testbed Authenticity: Evaluated against production open-source repositories (SymPy, Django, Sphinx, Pytest).
- Clean Patch Guarantee: 100% clean diffs passed through upstream test runner suites without regressions.
Research Provenance & Archives
The theoretical foundation and benchmark artifacts are archived across permanent scholarly repositories:
Air-Gapped Invariants & Zero Code Retention
For teams handling confidential intellectual property, defense software, or proprietary algorithms:
For security disclosures or technical collaboration inquiries, contact Tokenectomy Labs at daffa@kronumos.com.