Understanding Outcome Decomposition

OUTCOMES, NOT TASKS.

This is the most fundamental principle. The Product Tree contains OUTCOMES—states of the world that should exist—not tasks to be done.

ADRs and PDRs GOVERN, they don't implement.

• ADR = "when you do X, do it THIS way" — architectural constraints, no outcomes, no work, no tests
• PDR = "the product behaves THIS way" — product behavior constraints, no outcomes, no work, no tests
• NEED comes from stories/features/capabilities — they describe outcomes
• ADR compliance is verified through code review; PDR compliance through product validation

STORIES REQUIRED FOR IMPLEMENTATION.

A feature without stories is NOT ready for implementation—regardless of whether code already exists.

• Implementation STARTS at the story level, never at feature level
• Story-level tests provide atomic verification and property-based testing
• Feature-level tests alone are insufficient—they rarely cover edge cases
• Existing passing tests do not validate missing stories

WRONG: "Implementation exists and passes tests, so flat structure is fine"
RIGHT: "Feature needs stories before any implementation can be verified"

WRONG: "Feature-level tests are comprehensive enough"
RIGHT: "Only story-level tests enable atomic verification and property-based testing"

If you encounter a feature with implementation but no stories: recommend decomposition, not acceptance.

SHARED NEED → ENABLER:

When 2+ containers share a need, factor it into an ENABLER at the lowest convergence point:

54-component-library.capability/
├── 54-fixed-point-format.adr.md       # GOVERNS (no tests here)
├── 21-dsp-foundation.feature/         # ENABLER for shared need
│   └── 21-fixed-point-helpers.story/  # Implements shared helpers
│       └── tests/                     # Tests for the enabler
├── 54-dsp-cordic.feature/             # DEPENDS ON enabler
├── 76-dsp-nco.feature/                # DEPENDS ON enabler
└── 87-dsp-cic.feature/                # DEPENDS ON enabler

The ADR (and PDR) doesn't get tests. The ENABLER that satisfies the shared need gets tests.

The Product Tree replaces the backlog.

• Writing a spec = creating potential energy (a state that should exist)
• Passing tests = realizing potential (proving the state exists)
• The tree grows coherently—ideas must connect to existing structure

Structured Outcome Requirement:

Every outcome must be expressed as a typed, structured outcome (Scenario, Mapping, Conformance, or Property), have referenced test files, and pass agentic review. This is the quality gate — not syntactic constraints, but verified coherence between specs, tests, and passing tests.

Note: ADRs and PDRs don't have outcomes—they GOVERN. The requirement applies to outcomes in specs. See references/outcome-types.md for the four outcome types and their notations.

Principle 11:

Higher levels unaware of lower level breakdown. Features don't list story outcomes. Capabilities don't list feature outcomes. Completion bubbles up through test results, not spec references.

• /decomposing-prd-to-capabilities
• /decomposing-capability-to-features
• /decomposing-feature-to-stories

Key questions for each level:

Capability: What the Product CAN DO

A capability is a large, cross-cutting vertical slice of the product. It represents something the product is capable of doing, providing, or accomplishing.

Examples:

• "Document Generation"
• "Custody Account Management"
• "Monetization Platform"
• "Identity and Authentication Management"

Key characteristics:

• May be involved in one, many, or ALL customer journeys
• Represents a coherent area of product functionality
• Can start with a single feature and grow
• Up to 9 sibling capabilities (BSP numbering)

NOT a capability:

• A single API endpoint (too small → feature or story)
• A customer journey (journeys may span multiple capabilities)
• An implementation detail (belongs in story)

Feature: Significant Vertical Slice (≤7 Stories)

A feature is a significant and valuable vertical slice that can be implemented in at most 7 atomic stories.

Examples:

• "Stable Diffusion Generation"
• "User Auth" or more specifically "Password Auth" vs "Magic Link Auth"
• "User Lifecycle Management"
• "Export Documents"
• "Save Files to Cloud Storage"

The 7-story limit is critical:

IF feature needs > 7 stories
THEN split into multiple features

IF feature has only 1-2 stories
THEN it may be too small (or that's fine for now—it can grow)

Key characteristics:

• Implementable in at most 7 atomic stories
• Represents significant user value
• Can start with a single story and grow
• Up to 9 sibling features per capability

NOT a feature:

• A single function or method (too small → story)
• An entire product area (too large → capability)
• A technical task with no user value (may not belong in the tree at all)

Story: Atomic Implementation Unit

A story is the atomic unit of implementation—something that can be understood and implemented as a single coherent piece.

Examples:

• "Reset password"
• "Send reset password email"
• "Parse SPI configuration"
• "Generate Verilog for SPI master"

The level of abstraction depends on context:

• In a simple auth system: "Reset password" might be one story
• In a complex auth system: "Send reset password email" and "Validate reset token" might be separate stories

Key characteristics:

• Expressible as structured outcomes (Scenario, Mapping, Conformance, or Property)
• Atomic—can be implemented without partial states
• Has clear acceptance criteria
• Up to 9 sibling stories per feature (but aim for ≤7)

NOT a story:

• Multiple unrelated behaviors (split into separate stories)
• Vague requirements (not ready for engineering—needs discovery)
• Technical tasks without user-facing behavior (may be part of a story, not a separate one)

The Tree Grows Organically

Starting small is normal:

21-auth.capability/
└── 21-login.feature/
    └── 21-basic-login.story/

This is a valid structure. One capability, one feature, one story.

Growth happens at any level:

21-auth.capability/
├── 21-login.feature/
│   ├── 21-basic-login.story/
│   ├── 37-remember-me.story/        ← Story added
│   └── 54-login-throttling.story/   ← Story added
├── 37-registration.feature/          ← Feature added
│   └── 21-email-registration.story/
└── 54-password-reset.feature/        ← Feature added
    ├── 21-request-reset.story/
    └── 37-complete-reset.story/

BSP numbering enables insertion:

• Lower BSP = dependency (must be done first)
• Same BSP = parallel (can be done concurrently)
• Higher BSP = dependent (depends on lower numbers)

How to Decompose

Capability → Features:

Ask: "What significant vertical slices make up this capability?"

Each slice should be:

• Independently valuable
• Implementable in ≤7 stories
• A coherent unit of functionality

Feature → Stories:

Ask: "What atomic pieces make up this feature?"

Each piece should be:

• Expressible as a structured outcome
• Implementable as a single unit
• Clearly testable

The key question at every level:

Can this be expressed as a structured outcome (Scenario, Mapping, Conformance, or Property)? If not, decompose further or clarify requirements.

Common Mistakes

Mistake 1: Putting Story-Level Outcomes in Features

Wrong (the SPI example):

# Feature: Serial SPI

## Outcomes

### 1. SPI master transmits in mode 0 ← Should be a story

### 2. SPI modes 1-3 verified ← Should be a story

### 3. SPI slave responds ← Should be a story

### 4. Master-slave loopback works ← Feature-level OK

### 5. Lint-clean HDL ← Feature-level OK

Correct structure:

22-serial-spi.feature/
├── serial-spi.feature.md       # Outcomes 4 & 5 only (if needed)
├── 10-spi-master.story/        # Outcomes 1 & 2 become stories
├── 20-spi-slave.story/         # Outcome 3 becomes a story
└── ...

Why? Principle 11—features don't list story outcomes. Stories prove themselves through passing tests.

Mistake 2: Features with >7 Stories

If you need 12 stories, you have 2 features, not 1.

Mistake 3: Confusing Test Levels with Container Levels

Test levels (1, 2, 3) are about infrastructure needed:

• Level 1: No real infra (DI, temp dirs)
• Level 2: Real binaries/databases
• Level 3: Real services/credentials

Container levels (Story, Feature, Capability) are about scope of concern:

• Story: Atomic implementation
• Feature: ≤7 stories as a significant slice
• Capability: Cross-cutting product ability

These are orthogonal. A story can have Level 2 tests. A capability can have Level 1 tests.

Mistake 4: Removing All Outcomes from Features

WRONG: Agents removing ALL outcomes from features because "features don't list story outcomes."

Features MUST have their own outcomes. Principle 11 means features don't list outcomes that ARE stories (atomic implementation units). Features still have:

• Integration outcomes: Scenarios where multiple stories work together
• Quality gates: Cross-cutting verification (lint-clean, synthesizable)
• End-to-end scenarios: Complete feature workflows

# Feature: Serial SPI

## Outcomes

### 1. Master-slave loopback works ← KEEP (integration)

### 2. Generated HDL passes lint ← KEEP (quality gate)

What to MOVE to stories:

### 3. SPI master mode 0 works ← MOVE TO STORY (atomic)

Mistake 5: Treating ADRs/PDRs as Implementation

WRONG: "Where do we test the ADR?" or "ADR compliance story"

ADRs and PDRs GOVERN, they don't implement:

• ADRs/PDRs create NO work, NO outcomes, NO tests
• NEED comes from stories/features/capabilities
• ADR compliance verified through code review; PDR compliance through product validation

The ENABLER pattern for shared needs:

WRONG:
54-component-library.capability/
├── 54-fixed-point-format.adr.md
└── tests/
    └── test_adr_compliance.py    ← ADRs don't get tests!

RIGHT:
54-component-library.capability/
├── 54-fixed-point-format.adr.md       # GOVERNS (no tests)
├── 21-dsp-foundation.feature/         # ENABLER (has the NEED)
│   └── 21-fixed-point-helpers.story/  # Satisfies shared need
│       └── tests/                     # Tests verify outcomes
├── 54-dsp-cordic.feature/             # Uses enabler
└── 76-dsp-nco.feature/                # Uses enabler

Mistake 6: Assuming Existing Implementation Validates Missing Stories

WRONG reasoning:

"Implementation exists and is functional"
"Tests are comprehensive"
"Structure is flat - all tests at feature level"
"This may be intentional for a completed feature"

Why this is wrong:

1. Existence of code doesn't validate structure - The Outcome Engineering framework requires stories regardless of implementation state
2. Feature-level tests are never "comprehensive enough" - They rarely do property-based testing, edge case coverage, or atomic verification
3. "Flat structure works" is backlog thinking - Specs are permanent documentation, not work tracking
4. Completed ≠ correctly structured - Legacy code often lacks proper decomposition

RIGHT response when finding implementation without stories:

"Feature has implementation but no stories. This violates Outcome Engineering framework.
Recommend decomposing into stories to enable:
- Atomic verification of each behavior
- Property-based testing at story level
- Clear specification of edge cases
- Independent verification of each unit"

The test: If a feature has tests/ but no *.story/ directories, it needs decomposition—regardless of whether tests pass.

What Goes Where (Summary)

The Flow:

ADR/PDR constrains → Spec defines WHAT → Test PROVES → Ledger RECORDS

See references/what-goes-where.md for detailed taxonomy.

References

Skill mastery demonstrated when:

• Can explain the human interpretation of each level
• Can identify when a feature has too many stories (>7)
• Can recognize when story-level outcomes are incorrectly placed in features
• Understands that test levels and container levels are orthogonal
• Can decompose a capability into features, features into stories
• Knows that small trees (1-1-1) are valid starting points
• Recommends decomposition when finding features with implementation but no stories