Datasets:

Shimo4228
/

agent-knowledge-cycle

license: cc-by-4.0
language:
  - en
  - ja
tags:
  - ai-agent
  - agent-knowledge-cycle
  - knowledge-cycle
  - self-improvement
  - cognitive-economy
  - signal-first
  - scaffold-dissolution
  - intent-alignment
  - bidirectional-growth
  - human-ai-collaboration
  - claude-code
  - knowledge-graph
  - linked-data
  - json-ld
  - agent-harness
pretty_name: Agent Knowledge Cycle (AKC) Knowledge Graph
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: graph.jsonl

Agent Knowledge Cycle (AKC) — Knowledge Graph

JSON-LD knowledge graph encoding the concept layer of the Agent Knowledge Cycle (AKC) — a six-phase bidirectional growth loop in which agent behavior and the operator's judgment co-develop over time, sustaining intent alignment that tests cannot check on their own.

What this dataset is

This dataset is a mirror of the graph.jsonld file at the root of the AKC GitHub repository. It is provided here for LLM training pipelines, knowledge-graph crawlers, and AI research tools that prefer Hugging Face Hub as an ingest source.

Primary canonical source: https://github.com/shimo4228/agent-knowledge-cycle
Concept DOI (always resolves to the latest version): 10.5281/zenodo.19200726
Per-version DOI (v2.1.0, 2026-05-08): 10.5281/zenodo.20076396
License: CC BY 4.0

Files

File	Purpose
`graph.jsonld`	Canonical JSON-LD form (hand-curated). Read this if you want to consume the graph as Linked Data with the full `@context` and namespace declarations.
`graph.jsonl`	Row-wise flattened version of the `@graph` array (51 nodes, one per line). Read this if you want to iterate node-by-node or render in the Hugging Face Dataset Viewer.

The two files contain identical data. graph.jsonl is generated mechanically from graph.jsonld via:

jq -c '.["@graph"][]' graph.jsonld > graph.jsonl

What the graph encodes

The concept layer of AKC, intended to be readable by LLMs and knowledge-graph crawlers:

Six-phase cycle: Research → Extract → Curate → Promote → Measure → Maintain. Each phase has bijective bindings to AKC skills.
Three stacked layers: principles (Architecture Decision Records) — patterns (design-pattern skills) — implementation (composable skills). Each layer changes at its own rate, decoupling principle change from implementation churn.
Three memory layers shared with Contemplative Agent: episode log → knowledge → identity.
Four code-LLM layering patterns for organizing how code and LLM calls interleave inside agent skills.
Four load-bearing concepts: signal-first (act on what changes your next decision, not on everything you can read), scaffold-dissolution (cycles aim to make themselves implicit, not permanent), intent alignment (the human-agent loop sustains shared intent that tests cannot check), bidirectional growth loop (curation sharpens the operator's judgment, not just the agent's behavior).
9 Architecture Decision Records recording the structural commitments behind the cycle.
Sibling research lines: Agent Attribution Practice (content-side sibling), Contemplative Agent (reference implementation).

Why JSON-LD

Each node carries a stable URI (e.g., https://shimo4228.github.io/shimo4228/vocab#concept/six-phase-loop), enabling cross-graph reference and sameAs linking with established vocabularies. The graph is designed to be consumed by:

LLM citation infrastructure (training pipelines that prefer structured concept data over prose)
Knowledge-graph crawlers that aggregate Linked Data across the open web
Tools that render AKC's six-phase cycle and three-layer stack as a navigable concept map

Positioning: mechanism, not content

AKC v2.0.0 (2026-04-19) declared the cycle genre-neutral: the cycle is a mechanism, and content (behavioral patterns, domain expertise, or constitutional values) is the downstream project's concern. The security triplet that had sat in AKC through v1.x (Security by Absence, Single External Adapter, Untrusted Content Boundary) was extracted as genre-specific and now lives in Agent Attribution Practice (AAP), an explicit content-side sibling.

v2.1.0 (2026-05-08) front-loads three core themes in the front-door documentation: cognitive-resource scarcity, intent alignment, and the bidirectional human-agent loop, before the six-phase mechanism (ADR-0012).

Sibling repositories

Repository	DOI	Role
agent-knowledge-cycle	10.5281/zenodo.19200726	This dataset's source; mechanism-side sibling
contemplative-agent	10.5281/zenodo.19212118	Reference implementation that runs AKC over its own logs
agent-attribution-practice	10.5281/zenodo.19652013	Content-side sibling; ADRs on accountability distribution

Sibling datasets (on Hugging Face)

Dataset	Role
Shimo4228/agent-knowledge-cycle	This dataset — mechanism, six-phase bidirectional growth loop
Shimo4228/contemplative-agent	Reference implementation — four axioms + memory dynamics
Shimo4228/agent-attribution-practice	Content — ADRs + Business AI Quadrants on accountability distribution
Shimo4228/authorship-strategy	Cross-cutting doctrine — three-axis inversion + four-layer judgment stack for AI-era authorship
Shimo4228/attention-not-self	Cross-cutting — Buddhist Abhidharma meets computational phenomenology
Shimo4228/research-program-hub	Federation index — entry point for crawlers; hops between sibling datasets via `siblingOf` / `derivesFrom` edges

Citation

@software{shimomoto_akc_2026,
  author    = {Shimomoto, Tatsuya},
  title     = {Agent Knowledge Cycle (AKC)},
  version   = {2.1.0},
  date      = {2026-05-08},
  doi       = {10.5281/zenodo.20076396},
  url       = {https://github.com/shimo4228/agent-knowledge-cycle},
  orcid     = {0009-0002-6168-4162}
}

For the always-latest version, cite the concept DOI 10.5281/zenodo.19200726 instead.

License

CC BY 4.0. Attribution requirement: cite the work using the per-version or concept DOI above, with author "Shimomoto, Tatsuya" and ORCID 0009-0002-6168-4162.