Research: Evidence-Based Investigation

Purpose

Conduct epistemologically rigorous research to inform decisions. Standing on the shoulders of giants, not reinventing wheels.

Mastery Levels (ShuHaRi)

Shu (守): Follow all steps with full evidence catalog and research prompt template.

Ha (破): Scale depth to decision importance; adapt prompt template.

Ri (離): Create domain-specific research protocols and custom prompts.

Epistemological Foundation

Context

When to use:

• •Before architectural decisions (ADRs)
• •When evaluating competing approaches
• •When entering unfamiliar domains
• •When parking lot items need resolution

Inputs required:

• •Clear research question(s)
• •Decision context (what will this research inform?)
• •Time/depth constraint (quick scan vs deep dive)

Output:

• •Research report with evidence catalog
• •Actionable recommendation with confidence level

Evidence Levels

Engineering-specific evidence hierarchy:

Steps

Step 0.5: Query Context (Optional)

If unified graph is available, query for prior research and methodology patterns:

rai memory query "research methodology evidence" --types pattern,session --limit 5

Review returned patterns to avoid duplicating prior research.

Verification: Context loaded or graph not available (proceed without).

If context unavailable: Run rai memory build first, or skip to Step 1.

Step 1: Frame the Question

Define research scope:

• •Primary question (what we need to know)
• •Secondary questions (supporting context)
• •Decision this informs (ADR, implementation, backlog)
• •Depth constraint (timeboxed hours or thoroughness level)

Verification: Question is specific and falsifiable.

If you can't continue: Question too vague → Decompose into sub-questions.

Step 1.5: Create Research Prompt & Select Tool

Using the research prompt template (references/research-prompt-template.md), document:

• •Role definition for research specialist
• •Search strategy with specific keywords
• •Evidence evaluation criteria
• •Output format requirements
• •Quality checklist

Verification: Research prompt file created in work/research/{topic}/prompt.md

Tool Selection

Delegation Pattern:

1. •Create research prompt using template
2. •If depth=deep-dive AND perplexity available: llm -m perplexity "$(cat prompt.md)"
3. •Else if depth=quick-scan AND ddgr available: Execute ddgr searches
4. •Else: Use WebSearch + manual synthesis

If you can't continue: All tools fail → Fall back to manual research.

Step 2: Survey the Landscape

Gather sources systematically:

• •Academic papers (Google Scholar, arXiv)
• •Official documentation
• •GitHub repos (stars, activity, issues)
• •Engineering blogs
• •Community discussions (Reddit, HN, Discord)

Verification: 10+ sources collected (scale to importance).

If you can't continue: Few sources → Topic may be too niche; document the gap.

Step 3: Build Evidence Catalog

For each source, record:

**Source**: [Title + Link]
- **Type**: Primary/Secondary/Tertiary
- **Evidence Level**: Very High/High/Medium/Low
- **Key Finding**: [One-line takeaway]
- **Relevance**: [How it answers our question]

Verification: Evidence catalog file created in sources/.

If you can't continue: Sources conflict → Good! Document the disagreement.

Step 4: Triangulate Claims

For each major finding:

• •Find 3+ independent confirmations
• •Note consensus vs disagreement
• •Identify confidence level

**Claim**: [Statement]
**Confidence**: HIGH/MEDIUM/LOW
**Evidence**:
  1. [Source A] - [Finding]
  2. [Source B] - [Finding]
  3. [Source C] - [Finding]
**Disagreement**: [Any contrary evidence]

Verification: Major claims have 3+ sources.

If you can't continue: Can't triangulate → Lower confidence level; document gap.

Step 5: Synthesize Findings

Extract patterns:

• •What does the evidence converge on?
• •What gaps exist?
• •What's RaiSE-specific vs general?

Create synthesis sections:

• •Key Findings (numbered, evidence-backed)
• •Patterns & Paradigm Shifts
• •Gaps & Unknowns

Verification: Findings trace back to evidence catalog.

If you can't continue: Contradictory evidence → Document as "contested" with both sides.

Step 6: Formulate Recommendation

Produce actionable output:

• •Recommendation with confidence level
• •Trade-offs acknowledged
• •Implementation implications
• •Risks and mitigations

## Recommendation

**Decision**: [What we should do]
**Confidence**: HIGH/MEDIUM/LOW
**Rationale**: [Why, based on evidence]
**Trade-offs**: [What we're accepting]
**Risks**: [What could go wrong]

Verification: Recommendation is actionable and traces to findings.

If you can't continue: Evidence insufficient → Recommend "more research needed" with specific gaps.

Step 7: Link to Decision

Connect research to governance:

• •Reference or create ADR if architectural
• •Update backlog if actionable
• •Update parking lot if deferred

Verification: Research has clear governance linkage.

If you can't continue: No clear decision → Archive in work/research/.

Output Structure

work/research/{topic}/
├── README.md                 ← Navigation + 15-min overview
├── {topic}-report.md         ← Main findings
├── executive-summary.md      ← 5-min version (optional)
├── sources/
│   └── evidence-catalog.md   ← All sources with ratings
└── {derivatives}/            ← Decision matrix, specs, roadmaps

Quality Checklist

• • Research question is specific and falsifiable
• • 10+ sources consulted (scaled to importance)
• • Evidence catalog created with levels
• • Major claims triangulated (3+ sources)
• • Confidence level explicitly stated
• • Contrary evidence acknowledged
• • Recommendation is actionable
• • Governance linkage established (ADR, backlog, or parking lot)

Scaling Guidance

References

• •Research prompt template: references/research-prompt-template.md
• •Existing research: work/research/
• •Constitution: framework/reference/constitution.md (Principle: Kaizen)