Ken Imoto

• Building the LLMO Framework, LLMO Checker, and the Open LLMO Research Initiative
• Documenting AI-Native MEO as the local-business implementation of LLMO
• Applying Harness Engineering to products — dev, ops, and marketing with AI agents
• Publishing books & articles from hands-on practice

Updated July 2026

🌍 Rhythmic Monotonization in LLM-Generated Text Is Cross-Lingual: Evidence from Three Languages and Seven Models

Sequel to the two-layer fingerprint study: is rhythmic monotonization (lower burstiness, lower sentence-length CV) a Japanese quirk or a cross-lingual property? Seven LLMs against human baselines in English and Portuguese (853 Dev.to + 403 TabNews articles). AI lands on the monotonic side in all 70/70 cells (5 core metrics × 7 models × 2 languages), and the effect survives length residualization. Comma density is the lone metric whose sign flips by language (EN +0.45 / PT −0.83), motivating a three-layer signature/accent/dialect fingerprint framework. Current Claude-generation models monotonize at roughly 1/3 the strength of GPT-3.5 — the accent is fading.

🎼 Lexical and Rhythmic Fingerprints of Japanese LLM-Generated Text: A Cross-Model, Two-Layer Analysis

350 texts from 7 LLMs vs. 700 human articles, comparing lexical and rhythmic feature families under one controlled design. Rhythmic monotonization is real and model-universal (burstiness d = −0.96), yet vocabulary dominates human/AI discrimination (AUC 0.998 vs. 0.897), the opposite of the practitioner claim. Lexical profiles identify the generating model at 92% (rhythm: 51%), and a document-level re-analysis corrects our own prior study: 35% of the previously reported 651 excess words were token-counting artifacts.

⚡ When Free Executors Cost More: The Free-Executor Paradox in Iterative LLM Code-Repair Loops

Four LLM configurations on three Python code-repair tasks (breakage / refactor / feature-add) under a single deterministic correctness judge (mypy + ruff + pytest). The canonical "strong orchestrator + cheap executor" structure (Opus + local Qwen) turns out to be the most expensive cloud arm on every task because prompt-cached re-reads of executor summaries dominate any savings from delegating execution. Haiku-solo wins on dollar cost on the largest task (5.5× cheaper) at a 25% failure rate.

📊 Excess Vocabulary in Japanese AI-Generated Text: A Cross-Model Quantitative Analysis

350 samples from 7 LLMs vs. 977 human articles. 651 statistically significant excess words identified. Newer Claude generations show increasing vocabulary bias (TTR: 0.52→0.29). Cross-domain classifier achieves AUC=0.946 in-domain but fails completely across genres. First Japanese excess vocabulary study with coevolution evidence.

📝 AI Text Slop: Stylistic Convergence Across Six LLMs in Japanese Technical Writing

180 samples (6 models × 10 topics × 3 trials) measured with 16 pattern indicators. RLHF-aligned commercial models score significantly higher than OSS (Cohen's d = 1.01). Vocabulary and structural patterns dissociate ("Swallow Paradox"). Human Qiita articles score higher than AI on structural metrics, revealing cultural confounding.

🔵 AI Blue: Systematic Color Recognition Bias in Vision-Language Models

We tested 4 VLMs × 40 colors × 480 observations. Commercial models miss intermediate hues. 95% of AI-generated UI colors land on blue-purple. First quantitative link between VLM color bias and the "AI Slop" phenomenon.