Language: English | 日本語

Governance Mapping — Quadrants × ADRs × Controls

Purpose. Dense reference of which ADRs apply to each quadrant, what controls are required, and how redirect operates when something goes wrong. Use this alongside decision-tree.md after a quadrant has been identified.

At a glance

Quadrant	Governance level	Applicable ADRs	Required controls	Redirect target on failure
(1) Script	Light (out of scope for AAP)	—	code review, tests	code author / pipeline owner
(2) Algorithmic Search	Light (out of scope for AAP)	—	algorithm validation, search-space audit	algorithm author
(3) LLM Workflow	Medium	0001, 0003, 0004, 0005, 0006, 0007	scaffolding visibility, audit log, approval on behavior change, schema-conformant calls	role owner of the failing LLM call → routing-logic designer → upstream-data owner
(4) Autonomous Agentic Loop	High	all ten ADRs	+ pre-named gap-bearer (0008 + 0009), human approval on behavior-modifying writes (0005), full causal trace (0006), one accountable operator (0008), operation-phase placement requires a recorded Phase-crossing decision (0010)	the pre-named gap-bearer (principled gap; redirect at component level is foreclosed)

Phase note (orthogonal to Quadrant)

Phase (design vs operation) is independent of Quadrant — see README — Phase is a third dimension for the framing. The governance-specific consequence: placing Quadrant 4 in the operation phase invokes ADR-0010, adding the Phase-crossing decision (recorded at deployment time) to the controls listed above for Quadrant 4. Design-phase Quadrant 4 placements satisfy this automatically.

The empirical default for operation-phase workflows is a composition of Quadrant 1 + 3 (+ 2 where applicable). Deviations are surfaced through the decision tree Q0 + Q4 path.

Per-quadrant detail

(1) Script Quadrant — out of scope

LLM-related ADRs do not apply. Failure analysis routes to the code author or pipeline owner through normal SE channels.

(2) Algorithmic Search Quadrant — out of scope

Out of scope for this repository's LLM-focused ADRs. Failure routes to the algorithm author or the upstream specification owner.

(3) LLM Workflow Quadrant — Medium governance

See README — LLM Workflow Quadrant for the operational signal. Six ADRs apply:

ADR	What it requires for this quadrant
0001 Security by Absence	The harness does not ship code that executes shell commands or makes arbitrary outbound HTTP. Each LLM call is invoked through a narrow function with a documented schema.
0003 Untrusted Content Boundary	RAG content, accumulated state, and any content fetched from external sources is treated as untrusted when injected into prompts.
0004 Single External Adapter	The workflow as a deployed unit has at most one external adapter. Multi-surface workflows split into per-surface deployments.
0005 Human Approval Gate	Changes to the workflow's prompts, role definitions, schemas, or routing logic go through a human approval gate. Runtime behavior is not gated; the workflow itself is.
0006 Causal Traceability	Each LLM call's input, output, role identifier, and timestamp are logged. The workflow's control flow is reconstructible from logs.
0007 Scaffolding Visibility	Prompts, role definitions, schemas, routing rules, and the workflow's control flow are version-controlled files.

ADR-0008 (One Agent, One Human) is not strictly required in this quadrant — the workflow can be operated by a team if the audit chain is preserved — but if the workflow has external side effects, naming an accountable operator follows the same logic and is recommended.

ADR-0002 (Deterministic Prohibition at the Scaffolding Layer) and ADR-0009 (Triage Before Autonomy) are not in the table above because they apply across quadrants rather than within: 0002 is the fallback when 0001 cannot achieve absence; 0009 is the triage that selected this quadrant in the first place.

Redirect path for LLM Workflow Quadrant

When the workflow produces a wrong output, the operator walks the audit log to the failing component:

1. Which LLM call returned a wrong result? → role owner of that call (model selection, prompt design, schema specification).
2. Was the wrong call invoked due to a routing error? → routing-logic designer.
3. Was the call's input wrong? → upstream-data owner (data ingestion, RAG index curator, prior-step role owner).

Each step has a single owner. Distribution succeeds.

(4) Autonomous Agentic Loop Quadrant — High governance

See README — Autonomous Agentic Loop Quadrant for the operational signal and the attribution-gap commitment. All ten ADRs apply. ADRs 0005, 0006, 0008, 0009 carry the most weight because they are what makes the named gap-bearer's commitment operational. ADR-0010 adds the Phase-crossing decision when the placement is operation-phase.

ADR	What it requires for this quadrant
0001 Security by Absence	The agent does not ship capabilities it should not have. Especially load-bearing here because runtime tool selection means any capability can be invoked.
0002 Deterministic Prohibition at Scaffolding	Capabilities that cannot be made absent are gated by PreToolUse hooks or structural quarantine, outside the LLM.
0003 Untrusted Content Boundary	Episode logs, accumulated memory, and self-authored distillation are all treated as untrusted when read back into the loop.
0004 Single External Adapter	One agent process = one external adapter. Multi-surface deployments require splitting agents.
0005 Human Approval Gate	Load-bearing. Every behavior-modifying write — skills, rules, identity, constitution — goes through a named human, with an audit record for approved and rejected decisions alike. The gate is not a runtime "OK" click on every action; runtime side effects are bounded structurally by ADR-0001/0002/0004 instead.
0006 Causal Traceability	Load-bearing. The loop's iterations, tool calls, observations, and prompts are logged in append-only form. Causal trace must enable post-incident reconstruction even though component-level redirect is foreclosed.
0007 Scaffolding Visibility	The loop's prompt template, available tools, system prompt, and termination criteria are version-controlled files.
0008 One Agent, One Human	Load-bearing. Exactly one named human is bound to the agent process. Rotation is compatible if formal.
0009 Triage Before Autonomy	Load-bearing. The triage decision and the named gap-bearer are recorded at deployment time. Defaulting into this quadrant or naming the gap-bearer post-incident are forbidden.
0010 Phase Separation	Load-bearing for operation-phase placements. A Phase-crossing decision is recorded at deployment time: when a new pattern surfaces in operation, will the autonomous loop handle it dynamically in place, or route it back to design as feedback? Both answers admissible; "we'll figure it out" is not. Design-phase placements (coding agents, Deep Research) satisfy this automatically.

Redirect path for Autonomous Agentic Loop Quadrant

Component-level redirect does not succeed in this quadrant. The principled redirect impossibility is intrinsic: model output, tool selection, history reference, and prompt context blend at runtime, foreclosing post-hoc separability.

Redirect therefore operates at a different level:

1. System-level failure is identified through ADR-0006 causal traceability — the trace shows that the loop failed, even though it cannot show which judgment element contributed how much.
2. Responsibility lands on the pre-named gap-bearer (ADR-0008 + ADR-0009). The gap-bearer is the operator who absorbed the attribution gap as a pre-deployment commitment.
3. The gap-bearer decides the post-incident action — retrain, re-architect (move work to the LLM Workflow Quadrant), retire the deployment. Post-incident analysis cannot identify a single responsible component, so post-incident actions operate at the architectural level, not the component level.

This is a degraded form of redirect compared to (3) — it terminates at one party rather than distributing — but it is redirect, not the moral crumple zone failure where responsibility lands on whoever happens to be operationally proximate without having committed to bear it.

Cross-quadrant note: prohibition-strength hierarchy

ADRs 0001 → 0002 → 0003 form a prohibition-strength hierarchy. Quadrant-specific consequences:

• In the LLM Workflow Quadrant, scaffolding enforcement (0002) gates capabilities the workflow should not invoke even by mistake.
• In the Autonomous Agentic Loop Quadrant, scaffolding enforcement (0002) is essential because runtime tool selection means probabilistic prohibition will not hold under prompt injection.