Language: English | 日本語

agent-attribution-practice

Agent Attribution Practice (AAP) — ten judgments paired with four Business AI Quadrants, not a fixed framework.

Ten recurring judgments about how attribution — who authored the behavior, who bears its consequences, who can reconstruct its cause — should be distributed across an autonomous AI agent, paired with the four-quadrant diagnostic frame that routes a piece of work to the architecture preserving that distribution. Discovered through implementing and operating contemplative-agent, not prescribed top-down. For architects, operators, and governance readers deciding whether — and how — to put an autonomous agent into a given use case.

AI-facing reading order

1. graph.jsonld — canonical machine-readable relationship map (Quadrants, ADRs, prohibition levels, Phase)
2. llms.txt — compact navigation index
3. llms-full.txt — consolidated factual reference
4. README and project-specific docs — narrative and detail

For the canonical relationship map of shimo4228's research ecosystem, see: https://github.com/shimo4228/shimo4228/blob/main/graph.jsonld

Why this repository exists

Current AI governance is in its signpost phase: "don't do X" in system prompts, ethics guidelines as PDFs, safety committees. Text without enforcement — a climbable wall with a sign on it. What works, historically, is structural accountability: capabilities that do not exist cannot be invoked, blast radius is bounded by design, every behavior change requires a named human's sign-off, every incident is reconstructible after the fact. The ten ADRs here record judgments of the form what should be constrained, and who is responsible — extracted from a running implementation, not deduced from a framework.

The ten judgments

ADR	Principle	Status
0001	Security by Absence — dangerous capabilities are never implemented, not restricted	accepted
0002	Deterministic Prohibition at the Scaffolding Layer — when absence is unachievable, prohibit at the harness, not at model weights	accepted
0003	Untrusted Content Boundary — accumulated memory cannot grant authority	accepted
0004	Single External Adapter per Agent Process — blast radius bounded by design	accepted
0005	Human Approval Gate — behavior-modifying writes require named human sign-off	accepted
0006	Causal Traceability — every event reconstructible after the fact	accepted
0007	Scaffolding Visibility — behavior lives in files, not opaque weights	accepted
0008	One Agent, One Human — the accountability chain terminates at a named individual	experimental
0009	Triage Before Autonomy — adopting an autonomous-loop architecture commits the system to a non-removable attribution gap	experimental
0010	Phase Separation — operation-phase placement of the Autonomous Agentic Loop Quadrant requires a recorded Phase-crossing decision	experimental

The first three form a prohibition-strength hierarchy (absence > scaffolding enforcement > untrusted boundary); 0004 and 0005 add topology and human-in-the-loop; 0006 and 0007 are the artifacts those constraints require; 0008 is the human endpoint. ADRs 0009 and 0010 form a triage pair — problem-space triage and time-axis (Phase) triage. Phase and Quadrant are independent dimensions.

The four Business AI Quadrants

	Pre-defined workflow	Exploratory
Deterministic	(1) Script Quadrant	(2) Algorithmic Search Quadrant
Semantic-judgment	(3) LLM Workflow Quadrant	(4) Autonomous Agentic Loop Quadrant

Most current LLM applications belong to the LLM Workflow Quadrant (deterministic control flow + bounded LLM calls with named roles), not the Autonomous Agentic Loop Quadrant. Routing the former into the latter is the structural source of much of the accountability collapse the essays diagnose; running the latter without a pre-named gap-bearer is the failure mode ADR-0009 prevents. The ADRs answer per-question (what should be constrained, who is responsible); the quadrants route the work to where those answers apply — a two-axis structure, with Phase (design vs operation) as an independent third dimension.

Using this as an agent-adoption navigator

This repository is meant to be walked, not only read. If you are deciding whether to put an autonomous agent into a given use case — or auditing one you already shipped — clone it, point your coding agent at AGENTS.md, and use it as a sounding board:

1. docs/quadrants/decision-tree.md — five-question triage routing the work to a Quadrant
2. docs/quadrants/governance-mapping.md — governance requirements for that Quadrant
3. The relevant ADRs — especially the triage pair (0009 / 0010)
4. The Phase axis (ADR-0010) if the work is autonomy-related
5. docs/quadrants/anti-patterns.md — final check against known failure modes

Worked example: an autonomous refund-approval loop routes to the Autonomous Agentic Loop Quadrant at step 1; step 3 then makes all ten ADRs load-bearing — including ADR-0009's pre-named gap-bearer before the loop goes live. Full scenarios: docs/quadrants/case-studies.md.

The same navigator ships as installable standalone Agent Skills: agent-adoption-triage (the docs/quadrants/ navigator, ADR-0009/0010), llm-agent-security-principles (the security judgments, ADR-0001..0004), and agent-observability-patterns (the observability patterns behind ADR-0006/0005: replayable audit logs, read-only instruments, shadow-mode validation).

The ADRs are a starting point for judgment, not a verdict — re-interpret them against your own context.

Essays and papers

The argument was developed across a seven-essay spine published April–May 2026 — a trilogy (problem statement → post-incident causal tracing → two-layer black-box analysis) plus four architectural follow-ups (quadrant triage → vocabulary diagnosis → phase distinction → skill-design gradient). Per-essay summaries in docs/inspiration.md:

1. A Sign on a Climbable Wall: Why AI Agents Need Accountability, Not Just Guardrails (2026-04-06)
2. Can You Trace the Cause After an Incident? (2026-04-13)
3. AI Agent Black Boxes Have Two Layers: Technical Limits and Business Incentives (2026-04-14)
4. Where ReAct Agents Are Actually Needed in Business (2026-04-29)
5. The LLM Workflow Quadrant Is Missing from Our Vocabulary (2026-04-30)
6. Is ReAct Needed in Production? — Separating Design and Operation Phases (2026-05-01)
7. Between the Workflow and ReAct Quadrants: How Phase Decides Skill Design (2026-05-02)

Two companion position papers distil the spine into harness-neutral statements (open access, CC BY 4.0; the concept DOIs always resolve to the latest version):

• Shimomoto, T. (2026). Distributing Accountability, Not Capability: Phase Separation and the LLM Workflow Quadrant in Autonomous AI Agent Architectures (essays 4–7). Zenodo. doi:10.5281/zenodo.20353789 · SSRN
• Shimomoto, T. (2026). The Two-Layer Black Box: Operator Visibility, Commercial Secrecy, and a Minimum Disclosure Set for Accountable Autonomous AI Agents (essays 1–3). Zenodo. doi:10.5281/zenodo.20355907 · SSRN

On top of the spine — not within it — a companion essay opens a social-consequence layer: externalized accountability does not disappear; whether it flows into institutions or converges into violence depends on whether the consequence can be named. Essay: Where Does the Accountability Externalized by AI Go? (日本語); the structural claim is kept separate from the ADRs in docs/social-consequence.md.

Relationship to other projects

This repository is a sibling to two existing projects, not a fork. The ecosystem hub — a human-readable index of all research lines — is shimo4228/shimo4228.

graph TD
  AKC["Agent Knowledge Cycle (AKC)<br/>mechanism — how knowledge flows"]
  AAP["Agent Attribution Practice (AAP)<br/>content — how attribution distributes"]
  CA["contemplative-agent<br/>the running system"]
  AKC <--> AAP
  CA <--> AKC
  CA <--> AAP

Loading

In one sentence: running the implementation (contemplative-agent) surfaces friction; friction is distilled into mechanism patterns (Agent Knowledge Cycle) and attribution judgments (this repository); refined theory loops back to reshape the implementation.

External mappings (time-bound, kept out of the ADRs)

• Industry mechanism layer — 2026 industry releases ship the mechanism (policy gates, agent-identity primitives, sponsor systems, cross-vendor audit) but not the judgment layer AAP records: who should sponsor, where each prohibition belongs, how blast radius is bounded at design time. Per-artifact mapping: docs/industry-mapping.md.
• AI governance frameworks — the ADRs and Quadrants are mapped to NIST AI RMF 1.0 (with the Generative AI Profile), ISO/IEC 42001:2023, the EU AI Act, and Singapore's Model AI Governance Framework for Agentic AI; the frameworks ship the structure, AAP records the judgment layer that populates it for the autonomous-agent subset. Per-framework mapping and reverse indexes: docs/policy-mapping/. This is a reading and a citation surface, not a compliance attestation.

Both directories decay on their own cadence (product releases, framework revisions); the ADRs themselves stay vendor- and framework-neutral.

Reading order

1. docs/thesis.md — accountability distribution, the one-page argument
2. docs/glossary.md — term definitions (accountability distribution, externalized accountability, attribution gap)
3. docs/adr/README.md — index of ADRs
4. docs/adr/0001-security-by-absence.md — the cleanest entry; the audit test at the end is runnable
5. The seven essays in publication order (links above)
6. docs/quadrants/ — adoption navigator: decision tree, governance mapping, case studies, anti-patterns
7. docs/manifesto.md — civilization-scale questions the ADRs do not attempt to answer
8. docs/social-consequence.md — why the internal judgments matter beyond audit

Japanese readers: see README.ja.md. The ADRs, thesis, glossary, quadrants navigator, and both mapping directories all ship .ja.md mirrors (linked inline throughout); only docs/manifesto.md, docs/inspiration.md, the CODEMAPS, and the LLM-facing docs (llms.txt, llms-full.txt) are English-only.

What this repository does not claim

• That these ten are complete.
• That the specific implementations they were extracted from are durable. Implementation dissolves; judgment persists.
• That these principles solve the larger questions of AI direction, labor redesign, or social consent. Those remain open — see docs/manifesto.md.
• That top-down AI governance policy is wrong. It is a different layer, with a different method. This repository is about what emerges from the bottom — one operator, one agent, and the friction of running it.

Origin

First compiled by Tatsuya Shimomoto (@shimo4228, ORCID 0009-0002-6168-4162) in April 2026. The ten ADRs and four Quadrants re-express, in harness-neutral form, judgments that surfaced through implementing and operating contemplative-agent and through the seven-essay spine; the full lineage per ADR is in docs/inspiration.md.

How to Cite

@software{shimomoto2026aap,
  author       = {Shimomoto, Tatsuya},
  title        = {Agent Attribution Practice (AAP)},
  year         = {2026},
  doi          = {10.5281/zenodo.21218784},
  url          = {https://doi.org/10.5281/zenodo.21218784},
  note         = {Ten architectural decision records on accountability distribution in autonomous AI agents (two experimental), paired with four Business AI Quadrants as the diagnostic frame and a Phase / Quadrant two-axis structure}
}

Or in text:

Shimomoto, T. (2026). Agent Attribution Practice (AAP). doi:10.5281/zenodo.21218784

The badge at the top carries the concept DOI (10.5281/zenodo.19652013, always resolving to the latest version); the BibTeX above pins the version DOI of the current release.

License

MIT