Tokenectomy Labs

Autonomous cybernetic systems for automated program repair

Tokenectomy Labs is an independent AI systems and compiler research lab building a Dual-Brain Architecture for autonomous software engineering. We pair fine-tuned code language models with zero-allocation, deterministic Rust "Sub-Cortex" runtimes to deliver cost-bounded, verifiable program repair.

Flagship Engine: Kronumos Kairos

Kronumos Kairos is an open-weights program repair engine that pairs a 7B neural Cognitive Cortex (Qwen2.5-Coder-7B-Instruct) with the deterministic Tokenectomy Sub-Cortex.

Architectural invariants:

• Zero-allocation M2M Sub-Cortex: native Rust runtime doing AST traceback surgery, PikeVM/DFA secret sanitization, and deterministic POSIX diff re-anchoring, at under 5 ms latency.
• Zero Dirty Diff invariant: strict AST validation plus dry-run consensus gating. Under uncertainty or syntax invalidity, the system abstains, so no polluted patches reach the repository.
• Token efficiency: about 2,512 tokens per task on average, a 93.5% reduction versus multi-turn baselines.
• Zero API cost: runs on commodity/cloud GPUs (dual Tesla T4) with no commercial API dependency.

Benchmark Results

Evaluated on the full 500-instance SWE-bench Verified with Docker-based evaluation:

Research & Publications

• Preprint: Kronumos 2 Kairos: Cost-Bounded Automated Program Repair via Dual-Brain Cybernetic Sub-Cortex on SWE-bench Verified
• DOI: 10.5281/zenodo.22929676
• Code: Tokenectomy-Labs/Kronomus

Connect

• GitHub: Tokenectomy-Labs
• Lead Systems Architect: M N Daffa
• Research inquiries: daffa250r@gmail.com