Independent researcher working across AI alignment, multi-agent systems, efficient model computation, and measurement theory.

My research focuses on a recurring question:

When information or system states are transformed, compressed, updated, or passed between components, what remains reliable, what can fail, and how can we verify the difference?

The projects below address this question in different technical domains. They are independent research efforts, not claims of a single unified theory.

How can long-horizon AI systems preserve information integrity, authority boundaries, and reliable coordination?

• Boundary-Preserving Alignment
  A position paper on governed artifact promotion, bounded authority, provenance, human oversight, and the conditions under which AI-generated information may legitimately influence consequential actions.
  Paper (Zenodo)

• MRSI: Memory, Records, and Multi-Agent Coordination
  Four experimental papers and a technical synthesis examining provenance errors, memory repair, agreement quality, and approval boundaries in two- and three-role LLM systems.
  Collected Edition (Zenodo)

How can we reduce the computational cost of language models while understanding and controlling the resulting errors?

• NSN: Norm-Separated KV Cache Quantization
  Training-free low-bit KV-cache compression, including evaluations across model architectures, packed 3-bit caches for looped Transformers, and fused Triton execution.

• Norm-Separated Quantization
  Investigation of catastrophic INT4 quantization failures, separating magnitude and direction to identify and mitigate distinct failure mechanisms.

• LAB: Lyapunov-Aware Bit Allocation
  State-space model quantization using error persistence, output sensitivity, and dynamic range to allocate precision across recurrent states.

When a measurement, calibration, or model-derived score changes, under what conditions can its conclusions still be trusted?

• First-Order Probes
  Auditable sensor consistency checks and calibration updates under source shifts and contamination. Separates false-alarm evidence, fault-detection sensitivity, and decision coverage. Includes simulation code, Lean proofs, and original experimental data.

• Transformer Representation Geometry
  Empirical analysis of hidden-state geometry, principal-component concentration, and transferable feature readouts, including controls that challenge misleading geometric interpretations.

My work combines theoretical analysis, controlled experiments, implementation, and reproducible evaluation.

I aim to distinguish:

• Observed results from theoretical guarantees
• Task success from system-level correctness
• Model capability from operational authority
• Detection performance from the reliability of the resulting decisions
• Supported findings from unresolved or exploratory results

Where available, repositories include manuscripts, code, saved results, experimental protocols, and reproducibility instructions.

Negative results, limitations, and unresolved comparisons are part of the research record.

Independent Research | Japan · 2026