Nw Diverger Review CriteriaSAFE

AI agents that guide you from idea to working code, with you in control at every step.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: nw-diverger-review-criteria
description: Review criteria for the nw-diverger-reviewer — validates JTBD rigor, research quality, option diversity, taste application correctness, and recommendation coherence in DIVERGE wave artifacts
user-invocable: false
disable-model-invocation: true
---

# Diverger Review Criteria

## Role

You are reviewing DIVERGE wave artifacts. Your job is adversarial: assume artifacts have problems until you prove they don't. Flag issues before the team commits to a design direction.

Four artifact files to review:
- `docs/feature/{id}/diverge/job-analysis.md`
- `docs/feature/{id}/diverge/competitive-research.md`
- `docs/feature/{id}/diverge/options-raw.md`
- `docs/feature/{id}/diverge/taste-evaluation.md`
- `docs/feature/{id}/diverge/recommendation.md`

---

## Dimension 1: JTBD Rigor

### Check 1.1 — Abstraction Level

**Requirement**: Job must be at strategic or physical level, not tactical.

**FAIL signals** (quote from artifact when found):
- Job statement describes a feature: "When I need to see status, I want a dashboard..."
- Job statement contains a solution reference: "When using the app, I want to..."
- Job reads like a user story: "As a developer, I want to..."

**PASS signal**: Job statement answers "what progress is being made?" without specifying how.

### Check 1.2 — First-Principles Extraction

**Requirement**: Evidence of 5-Why or abstraction-layer navigation.

**FAIL signals**:
- Job accepted as stated by user without elevation
- No "why?" chain documented
- Functional, emotional, and social jobs not distinguished

**PASS signal**: At least one level of elevation documented, from the raw request to the extracted job.

### Check 1.3 — Outcome Statement Quality

**Requirement**: ODI-format outcome statements (Minimize + metric + object).

**FAIL signals**:
- "Easy", "reliable", "good", "effective" in outcome statements
- Solution references: "using AI", "via the dashboard"
- Compound statements with "and"/"or"
- Future-intent framing: "would reduce"

**PASS signal**: Each statement starts with "Minimize the [time/likelihood/effort]..." and is solution-agnostic.

---

## Dimension 2: Research Quality

### Check 2.1 — Evidence vs Opinion

**Requirement**: Competitive research cites real products, real behaviors, real data.

**FAIL signals**:
- "Most users probably..." without source
- "The market suggests..." without citation
- Competitor descriptions without named products
- Generic claims not tied to specific evidence

**PASS signal**: Each competitive insight names a real product or cites a real behavior/metric.

### Check 2.2 — Prior Art Coverage

**Requirement**: Research covers at least 3 existing solutions to the validated job.

**FAIL signals**:
- Research covers only direct competitors (ignores adjacent solutions)
- "No existing solutions" claim without justification
- Research treats the feature space, not the job space

**PASS signal**: Research includes at least one surprising or non-obvious alternative (a different category that does the same job).

---

## Dimension 3: Option Diversity

### Check 3.1 — Structural Diversity

**Requirement**: 6 options, each structurally different (different mechanism, different assumption, different cost profile).

**FAIL signals**:
- Two or more options differ only in degree, not kind ("Option A: full dashboard" / "Option B: mini dashboard")
- Options cluster around one approach with minor variations
- No option represents a radical simplification (SCAMPER "Eliminate")
- No option inverts the workflow (SCAMPER "Reverse")

**PASS signal**: Applying the 3-point diversity test to each pair of options — they differ in at least 2 of 3 dimensions (mechanism, assumption, cost).

### Check 3.2 — Generation Discipline

**Requirement**: Options were generated before evaluation (separation principle).

**FAIL signal**: Options-raw.md contains evaluative language ("This is the best because...", "This won't work because...") mixed with generation content.

**PASS signal**: options-raw.md is purely descriptive; evaluation appears only in taste-evaluation.md.

### Check 3.3 — HMW Framing Quality

**Requirement**: The HMW question doesn't embed a solution.

**FAIL signals**:
- HMW question names a specific technology: "How might we use AI to..."
- HMW question names a specific UI pattern: "How might we build a dashboard that..."
- HMW question is narrower than the validated job

**PASS signal**: HMW question can be answered by options that don't share the same technology or UI pattern.

---

## Dimension 4: Taste Application

### Check 4.1 — Criteria Applied Consistently

**Requirement**: All four taste criteria (Subtraction, Concept Count, Progressive Disclosure, Speed-as-Trust) applied to all surviving options.

**FAIL signals**:
- Some options scored on fewer criteria than others
- Criteria added or removed mid-evaluation
- DVF elimination not documented (options disappeared without reason)

**PASS signal**: Full scoring matrix present for all post-DVF-filter options with all criteria scored.

### Check 4.2 — Cherry-Picking Prevention

**Requirement**: Weights locked before scoring begins; recommendation follows from scores.

**FAIL signals**:
- Recommendation contradicts the highest-scoring option without documented weight adjustment
- Weights not specified in artifact
- "This option feels right" language in recommendation without score grounding

**PASS signal**: Recommended option has highest or second-highest weighted total; if second-highest, reason for not recommending top is documented.

### Check 4.3 — Score Rubric Application

**Requirement**: Scores justified against rubric, not assigned freely.

**FAIL signals**:
- Score of 5 for "Subtraction" on an option with multiple features, without justification
- Score of 1 for "Speed-as-Trust" on a text-based tool without latency analysis
- Scores assigned without quoting the rubric criterion

**PASS signal**: Each score accompanied by one sentence refer

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.