Language: English | 日本語

Decision Tree — Routing a Piece of Work to a Quadrant

Purpose. Convert ADR-0009 Triage Before Autonomy and ADR-0010 Phase Separation into a five-question navigator usable in design reviews. Each branch references the ADRs that become load-bearing once the quadrant and phase are chosen, so this is a judgment chain navigator, not a fixed framework.

The five questions

Q1–Q4 are the problem-space triage from ADR-0009. Q0 surfaces the Phase-crossing decision from ADR-0010 when the work is operation-phase and the result of Q1–Q4 is the Autonomous Agentic Loop Quadrant.

Phase and Quadrant are independent dimensions (see README — Phase is a third dimension). Q0 is not a Quadrant filter; it is a prompt to surface the Phase-crossing decision when it is load-bearing.

Q0. Is this work in the design phase or the operation phase?

• Design phase (exploratory work, workflow-being-built, prototyping, unknown structure being analyzed) → continue to Q1–Q4. All Quadrants are legitimate candidates depending on the work. Coding agents and Deep Research tools are typically here as design-phase residents (each session is an independent design session).
• Operation phase (deployed workflow running on its known path, predictability is the requirement) → continue to Q1–Q4. All Quadrants are still candidates depending on the work, but if Q1–Q4 yields the Autonomous Agentic Loop Quadrant, the Phase-crossing decision must be recorded explicitly (see Q4 Part B below). The empirical default for operation-phase workflows is Quadrant 1 + 3 (+ 2 where applicable), but this is a default, not a requirement.

The Phase decision is itself recorded alongside the deployment. "This is operation phase, default composition (Quadrant 1 + 3)" is a complete record. "This is operation phase, Quadrant 4 admitted, Phase-crossing decision: route new patterns back to design" is a complete record. "We didn't decide" is not.

The four problem-space questions

┌─────────────────────────────────────────────────────────────┐
│ 1. Can the work be expressed as deterministic rules?        │
│    (no semantic judgment required at any step)              │
└─────────────────────────────────────────────────────────────┘
            │
        YES │                                    NO
            │                                     │
            ▼                                     ▼
┌─────────────────────────────────┐   ┌─────────────────────────────────┐
│ 2. Is the search space          │   │ 3. Can the workflow             │
│    static and finite?           │   │    (the role of each LLM call)  │
│                                 │   │    be defined in advance?       │
└─────────────────────────────────┘   └─────────────────────────────────┘
        │             │                      │             │
    YES │         NO  │                  YES │         NO  │
        │             │                      │             │
        ▼             ▼                      ▼             ▼
   ┌────────┐    ┌────────────┐         ┌─────────┐    ┌─────────────────┐
   │ Script │    │ Algorithmic│         │   LLM   │    │ 4. Can a        │
   │Quadrant│    │   Search   │         │ Workflow│    │    pre-named    │
   │        │    │  Quadrant  │         │ Quadrant│    │   gap-bearer    │
   │        │    │            │         │         │    │  be identified? │
   └────────┘    └────────────┘         └─────────┘    └─────────────────┘
                                                              │
                                                          YES │      NO
                                                              ▼      │
                                                       ┌────────────┐│
                                                       │ Autonomous ││
                                                       │  Agentic   ││
                                                       │Loop Quadrant││
                                                       └────────────┘│
                                                                     ▼
                                                          ┌────────────────┐
                                                          │ Decision: do   │
                                                          │ not deploy     │
                                                          │ (gap unmanaged)│
                                                          └────────────────┘

Question-by-question detail

Q1. Can the work be expressed as deterministic rules?

If the work can be specified completely with rules — "given this input shape, produce this output by these steps, no semantic judgment required" — the answer is yes. Form filling, data validation, threshold checks, schema transformations, lookup-and-format work.

• Yes → continue to Q2.
• No → continue to Q3 (an LLM is in the loop somewhere).

The test for "deterministic": can two reasonable engineers produce identical outputs without consulting each other, given the same input and the same specification? If yes, no semantic judgment is required.

Q2. Is the search space static and finite?

For deterministic work, distinguish enumerable tasks from exploratory tasks.

• Yes → Script Quadrant. The ten ADRs are largely out of scope.
• No → Algorithmic Search Quadrant. Out of scope for AAP's LLM-focused ADRs; standard SE accountability applies.

Q3. Can the workflow be defined in advance?

For LLM-involved work, this is the central triage question. The "workflow" here means: the role of each LLM call, the order in which calls are made, and the data flow between them.

• Yes → LLM Workflow Quadrant. The default for most current LLM applications. Apply ADRs 0001, 0003, 0004, 0005, 0006, 0007. See governance-mapping.md for what each ADR requires for this quadrant.
• No → continue to Q4.

The test for "workflow definable in advance": can you draw the sequence of LLM calls on a whiteboard before running the system, labeling each call's role and schema? If yes, the workflow is pre-defined. If the next call's role depends on what the previous call's output happened to be, the workflow is exploratory.

Q4. Can a pre-named gap-bearer be identified, and (if Q0=operation) is the Phase-crossing decision recorded?

Reaching this question means the work genuinely requires the Autonomous Agentic Loop Quadrant. The attribution gap is intrinsic; redirect cannot succeed at the level of separable contributions.

The question has two parts; both must be answered yes for deployment.

Part A (always required). Who absorbs the attribution gap? ADR-0009 forbids leaving this to post-incident discovery (the moral crumple zone failure mode).

Part B (only if Q0 = operation phase). Is the Phase-crossing decision recorded — i.e., when a new pattern surfaces in operation, will it be handled by the autonomous loop in place, or routed back to the design phase as feedback? ADR-0010 forbids implicit Phase-crossing decisions for operation-phase Quadrant-4 placements. For design phase placements, this part is automatically satisfied (the design phase is where Phase-crossing happens).

• Yes (both parts) → Autonomous Agentic Loop Quadrant. Apply all ten ADRs (see governance-mapping.md for the per-ADR breakdown; ADRs 0005, 0006, 0008, 0009, 0010 carry the most weight). The gap-bearer (and Phase-crossing decision, when applicable) are recorded alongside the architectural choice at deployment time.
• No → Decision: do not deploy. The configuration the ADR set is designed to prevent — Autonomous Agentic Loop Quadrant work running either without a gap-bearer (Part A failure) or without an explicit Phase-crossing decision in operation (Part B failure).

The test for "gap-bearer identifiable": can the deployment record name a specific human (per ADR-0008), an organizationally identifiable role with formal succession, or a contractual party (insurance pool, regulatory body) who acknowledges, at deployment time, that they bear responsibility for failures of this work? If none of these can be named, the answer is no.

The test for "Phase-crossing decision recorded": can the deployment record state, in one sentence, what happens to a new pattern that surfaces during operation? "Route to design as feedback" and "Handle in the autonomous loop in place" are both admissible answers. "We'll figure that out when it happens" is not.

Hybrid architectures

Real systems often combine quadrants. ADR-0009 permits hybrids when the boundary between quadrants is itself a structural choice and the Autonomous Agentic Loop portion is treated as such (with a named gap-bearer for that portion).

Common shapes:

• Plan-and-Execute — autonomous loop generates the plan; LLM Workflow executes each step. The autonomous portion is the planning phase.
• Router agent — LLM Workflow with a routing call that classifies inputs into branches. The router itself is bounded; the branches are bounded.
• Tiered handoff — LLM Workflow handles common cases; escalates to (approval-gated) Autonomous Agentic Loop for novel cases.

For each hybrid, run the decision tree on each portion separately. The autonomous portion still needs Q4 answered.

When to re-run the triage

The triage is a design-time decision, but the answer can change as the work matures:

• Phase transitions. Work that begins in the design phase often matures into operation as the path becomes understood. Re-run Q0 when this happens; the resulting Quadrant placement may shift from 4 to 3 (a legitimate maturation, not a regression).
• Quadrant transitions. Tasks that start in the Autonomous Agentic Loop Quadrant often move to the LLM Workflow Quadrant once the path is understood. Re-run Q1–Q4.
• Vendor capability shifts. A capability that was formerly out of reach (Q1=No, Q3=No) may become workflow-able as models and tooling improve, moving the work from Quadrant 4 to Quadrant 3.
• Phase-crossing decision shifts. An operation-phase deployment that had recorded "route new patterns back to design" may, after experience, decide to switch to "handle in the autonomous loop in place" — or vice versa. Re-record the Phase-crossing decision alongside the architectural change.
• Regulatory shifts. A regulatory change can change the answer to Q4 (whether a gap-bearer can be named for a given domain).

When in doubt, re-run the five questions and document the new answers alongside the architectural change.