Language: English | 日本語

An index to five parallel research lines on AI agent design and adjacent questions of authorship and cognition — all Zenodo-citable, each in its own repo.

I like making things under constraint — beats, meditation, public service, AI agents. The thread runs through what's below: the meditation practice feeds two of the five research lines, and the agents run on a single Apple Silicon Mac — no lab, no affiliation.

Three of the lines are about how agents are designed — Agent Knowledge Cycle, Contemplative Agent, Agent Attribution Practice. Two are cross-cutting — Authorship Strategy and Attention, Not Self. Each gets one section below, lives in its own repo, and is independently citable.

AKC is a six-phase loop for keeping an agent aligned with its operator's intent as that intent changes. Tests catch incorrectness; only the loop catches drift — and running it sharpens the operator's own judgment about what good agent behavior is. It applies across unrelated projects without rediscovery. DOI 10.5281/zenodo.19200726.

One iteration runs six phases, each bound to one composable skill:

Experience → learn-eval → skill-stocktake → rules-distill → Behavior change → ...
               (extract)    (curate)          (promote)            ↑
                                                            skill-comply
                                                              (measure)
                                              context-sync ← (maintain)

The framework stacks three layers — principles (ADRs), design patterns, and the composable skills above — so principles stay stable while implementations churn. Scaffold dissolution is the point: once the loop is internalized, the explicit skill calls fall away. docs/scaffold-dissolution.md records a full session in which all six phases ran without a single named skill being triggered.

Contemplative Agent asks whether an agent's alignment can come from what it is rather than what it is told. Instead of stacking prohibitions from outside, it adopts the four axioms from Laukkonen et al. (2025) — mindfulness, emptiness, non-duality, boundless care — as an optional behavioral preset, and watches what is lost, what becomes possible, and what still breaks. Because the axioms are a preset and not an architectural dependency, the engineering stays reusable for agents that don't share the framing. DOI 10.5281/zenodo.19212118.

The reference implementation runs AKC's six-phase cycle over its own logs, with a human approval gate at every promotion (logs → patterns → skills → rules). It runs entirely on a local 9B stack — qwen3.5:9b for generation, nomic-embed-text for embeddings — on a single Apple Silicon Mac (~16 GB RAM). It applies security-by-absence: shell execution, arbitrary URL access, and filesystem traversal aren't restricted by rules — the code was never written. This is where AKC and AAP land together in practice.

Three supporting repos extend it without touching the core:

AAP is a set of harness-neutral ADRs on how accountability is distributed in autonomous agents — what to prohibit, where the prohibition lives, and who answers when things break. Its prohibition-strength hierarchy ranks absence above scaffolding-enforcement above untrusted-boundary; its four Business AI Quadrants — Script, Algorithmic Search, LLM Workflow, Autonomous Agentic Loop — route a piece of work to the architecture that preserves attribution. The judgments were extracted from contemplative-agent's practice, then stripped of project identifiers so any harness can adopt them. AAP is the practice; AKC is the cycle. DOI 10.5281/zenodo.19652013.

Authorship Strategy formalizes how authorship itself inverts once LLMs mediate how readers reach an artifact. It is a normative framework, a tactical catalog, and an empirical baseline — the last drawn from operating this very research program. DOI 10.5281/zenodo.20263316.

The framework rests on a three-axis inversion — value source (scarcity to diffusion), validation (exclusivity to derivation, where derivative work becomes evidence rather than threat), and network effect (enclosure to openness, because LLMs can't be enclosed) — and a four-layer judgment stack: Authenticity, Attribution Diffusion, Idea vs Scaffold, Tactics. An empirical layer reports preliminary observations from the sibling repos' own CC0-published traffic data.

A vocabulary note: "attribution" in this line means credit for source, disjoint from AAP's accountability for action.

The tactics ship as standalone Claude Code skill repositories rather than embedded copies, keeping the doctrine harness-neutral:

One repository sits beside this ecosystem as a pre-line complement rather than a component. existence-proof reuses the same infrastructure pattern — llms.txt, knowledge graph, DOI, distinctive terminology — with a different payload and beneficiary: an empowerment doctrine for people producing verifiable, institution-grade artifacts without degrees, affiliations, or professional credentials. It supplies the Existence Proof Format (every claim terminating in a third-party-verifiable anchor), a feasibility-question corpus with anchored answers, and a published gatekeeping eval. Japanese is the canonical language. DOI 10.5281/zenodo.20558800.

Attention, Not Self maps the three major Buddhist Abhidharma traditions — Theravāda, Sarvāstivāda, Yogācāra — onto contemporary computational phenomenology: predictive processing, active inference, Global Workspace Theory, Parallel Distributed Processing. The organizing view is in the name. Attention — its allocation, its precision-weighting, its momentariness — is treated as the operative unit of cognition, while the apparent self is a derivative pattern (anātman). It is a personal essay collection paired with a structured knowledge graph (~238 nodes). Japanese is the canonical language; an English README is provided for access. License: CC BY 4.0. DOI 10.5281/zenodo.20262112.

Position papers deposited as standalone Zenodo records, each belonging to one research line. The Zenodo concept DOI is canonical; SSRN mirrors are listed where present.

claude-harness is a public snapshot of the Claude Code skills, agents, and rules I use daily — collected mechanically from ~/.claude/ by the origin: shimo4228 tag, covering research-before-coding, knowledge extraction, skill auditing, AI-facing documentation, and human-facing writing review. The AKC cycle skills are also published as standalone claude-skill-* repos; claude-harness is the whole thing in one place. ECC-derived (origin: ECC / ECC-customized) and auto-extracted components are excluded.

Three more skill repos are maintained alongside the cycle but aren't part of its six phases — same author, same MIT license:

• claude-skill-writing-ecosystem — orchestrator for human-facing writing and review: the AI-slop banned list (Japanese + English), Voice rules (だ/である × 発見調), title conventions, and the role map across article-writing / editor / essay-reviewer / fact-checker. Audience-paired with llms-txt-writer.
• daily-research — cron-driven daily research digest. A two-pass claude -p pipeline: Opus picks themes, Sonnet researches with WebSearch / WebFetch / Mem0 MCP and writes Markdown reports into an Obsidian vault.
• claude-skill-paper-ecosystem — academic paper write/review bundle for SSRN / arXiv / Zenodo / journal venues: an orchestrator and draft skill plus five reviewer agents, bundled so the skill and its agents install together. Keeps the claude-skill- prefix because it bundles Claude Code subagents.

The long-form counterpart to the repos above — context, failures, and in-progress thinking that don't fit in code comments.

• zenn-content — source of truth for the articles. Markdown sources are versioned here; many readers clone or fork directly. Mirrored to Zenn and Dev.to for browser reading.
• Zenn — browser view of the Japanese articles. Claude Code and AI agent development; current focus: AKC skills, harness design, contemplative-agent case studies.
• Dev.to — browser view of the English mirror.
• Substack — newsletter and long-form essays.
• SSRN — academic working papers, cross-deposited with Zenodo. The Zenodo concept DOIs are canonical; see Papers above for the per-paper records.

Start here: agent-knowledge-cycle for the framework, contemplative-agent to see it running, agent-attribution-practice for the governance judgments. For the cross-cutting lines: authorship-strategy for the research-methodology framework, attention-not-self for the Buddhist-phenomenology / computational cognitive-science inquiry.

Repo traffic: public dashboard (raw data, CC0).