---

SKILL-SPECIFIC GUIDANCE

When to Use This Skill

• Complete research lifecycle from literature review to production (Pipelines A-I)
• Multi-month academic projects requiring 3 quality gates
• NeurIPS/ICML/CVPR submissions with reproducibility requirements
• Research requiring systematic methodology (PRISMA, ACM badging)
• Coordinating 9 pipelines with 15+ specialized agents

When NOT to Use This Skill

• Quick investigations (<1 week, use researcher skill)
• Single-pipeline workflows (use specific skills)
• Industry projects without academic rigor
• Prototyping without publication goals

Success Criteria

• All 3 Quality Gates passed (Foundations, Development, Production)
• Minimum 50 papers reviewed (Pipeline A)
• Baseline replicated within +/- 1% (Pipeline D)
• Novel method validated (p < 0.05, d >= 0.5)
• Holistic evaluation across 6+ dimensions
• Reproducibility package tested in fresh environments
• Ethics review completed (data bias audit, fairness metrics)

Edge Cases & Limitations

• Gate 1 failure: incomplete literature review, missing SOTA benchmarks
• Gate 2 failure: insufficient ablations, statistical power too low
• Gate 3 failure: production infrastructure not validated, monitoring gaps
• Multi-modal data: expand holistic evaluation to modality-specific metrics
• Limited compute: prioritize smaller ablation sets, document constraints

Critical Guardrails

• NEVER skip Quality Gates (use gate-validation for rigorous checks)
• ALWAYS document full pipeline execution (A through I, no shortcuts)
• NEVER claim production readiness without Gate 3 validation
• ALWAYS coordinate ethics review (ethics-agent) before Gate 1
• NEVER bypass reproducibility requirements (archivist agent mandatory)

Evidence-Based Validation

• Validate Gate 1: verify 50+ papers, SOTA benchmarks, research gaps
• Validate Gate 2: confirm 5+ ablations, p < 0.05, effect size d >= 0.5
• Validate Gate 3: test production deployment, monitoring, rollback strategies
• Cross-validate pipelines: ensure Pipeline D baseline feeds into Pipeline E
• Verify agent coordination: check memory-mcp state, confirm handoffs logged

---

name: deep-research-orchestrator description: Meta-orchestrator for complete Deep Research SOP lifecycle managing 3 phases, 9 pipelines (A-I), and 3 quality gates. Use when starting new research projects, conducting systematic ML research, or ensuring rigorous scientific methodology from literature review through production deployment. Coordinates all SOP skills and agents for end-to-end research execution. version: 1.0.0 category: research tags:

• research
• analysis
• planning author: ruv

---

Deep Research Orchestrator

Master orchestration skill for the complete Deep Research Standard Operating Procedure (SOP), managing the entire research lifecycle from ideation through production deployment with rigorous quality gates.

Overview

Purpose: Orchestrate complete research lifecycle following Deep Research SOP methodology

When to Use:

• Starting new machine learning research projects
• Conducting systematic scientific research with reproducibility requirements
• Academic paper submission with artifact evaluation
• Production ML deployment requiring ethics, fairness, and safety validation
• Research requiring regulatory compliance (FDA, EU AI Act)
• Multi-month research projects with team coordination

Quality Gates: Manages ALL 3 quality gates (Data & Methods, Model & Evaluation, Production & Artifacts)

Prerequisites:

• Research question formulated
• Resources allocated (compute, datasets, team)
• Institutional approvals (IRB if needed)
• Memory MCP configured for cross-session persistence

Outputs:

• Complete research artifact package
• Published paper with reproducibility artifacts
• Production-ready model with ethics validation
• All quality gate checklists (3 gates APPROVED)
• Comprehensive documentation (datasheets, model cards, method cards)
• DOI-assigned artifacts (Zenodo, HuggingFace)

Time Estimate: 2-6 months (full research lifecycle)

• Phase 1 (Foundations): 2-4 weeks
• Phase 2 (Development): 6-12 weeks
• Phase 3 (Production): 2-4 weeks

Skills Orchestrated: baseline-replication, method-development, holistic-evaluation, literature-synthesis, reproducibility-audit, deployment-readiness, research-publication, gate-validation

Agents Used: ALL 4 P0 agents (data-steward, ethics-agent, archivist, evaluator) + system-architect, coder, tester, reviewer, researcher

---

Quick Start

1. Initialize Research Project

# Initialize Deep Research SOP project
npx claude-flow@alpha hooks pre-task \
  --description "Deep Research SOP: [Project Name]"

# Create project structure
mkdir -p deep-research-project/{phase1-foundations,phase2-development,phase3-production,gates,docs}

# Store research question in memory
npx claude-flow@alpha memory store \
  --key "sop/project/research-question" \
  --value "How does multi-scale attention improve long-range dependency modeling in vision transformers?"

2. Run Phase 1 (Foundations) - Quality Gate 1

# Literature synthesis
claude-code invoke-skill literature-synthesis

# Dataset validation
npx claude-flow@alpha sparc run data-steward "/init-datasheet"

# Baseline replication
claude-code invoke-skill baseline-replication

# Ethics review (initial)
npx claude-flow@alpha sparc run ethics-agent "/assess-risks --component dataset --gate 1"

# Gate 1 validation
claude-code invoke-skill gate-validation --gate 1

3. Run Phase 2 (Development) - Quality Gate 2

# Method development
claude-code invoke-skill method-development

# Holistic evaluation
claude-code invoke-skill holistic-evaluation

# Ethics review (model)
npx claude-flow@alpha sparc run ethics-agent "/assess-risks --component model --gate 2"

# Gate 2 validation
claude-code invoke-skill gate-validation --gate 2

4. Run Phase 3 (Production) - Quality Gate 3

# Reproducibility audit
claude-code invoke-skill reproducibility-audit

# Deployment readiness
claude-code invoke-skill deployment-readiness

# Archival
npx claude-flow@alpha sparc run archivist "/init-model-card"

# Gate 3 validation
claude-code invoke-skill gate-validation --gate 3

5. Publication

# Research publication
claude-code invoke-skill research-publication

---

Detailed Instructions

PHASE 1: FOUNDATIONS (2-4 weeks)

Pipeline A: Literature Synthesis

Objective: Systematic literature review identifying SOTA methods, gaps, opportunities

Execution:

claude-code invoke-skill literature-synthesis \
  --query "vision transformers attention mechanisms" \
  --databases "arxiv,semantic-scholar,papers-with-code" \
  --output phase1-foundations/literature/

Deliverables:

• Literature review document (50-100 papers)
• SOTA performance benchmarks
• Research gap analysis
• Hypothesis formulation

Agent: researcher

---

Pipeline B: Data & Ethics Foundation

Objective: Dataset validation, bias audit, ethics clearance

Execution:

# Data steward: Create datasheet
npx claude-flow@alpha sparc run data-steward \
  "/init-datasheet --dataset ImageNet --output phase1-foundations/datasheet.md"

# Data steward: Bias audit
npx claude-flow@alpha sparc run data-steward \
  "Run bias audit on ImageNet dataset following Gebru et al. 2021"

# Ethics agent: Risk assessment
npx claude-flow@alpha sparc run ethics-agent \
  "/assess-risks --component dataset --gate 1"

Deliverables:

• Datasheet (Form F-P1)
• Bias audit report
• Ethics review (Gate 1)
• IRB approval (if human subjects)

Agents: data-steward, ethics-agent

---

Pipeline C: PRISMA Protocol (if systematic review)

Objective: PRISMA-compliant systematic literature review

Execution:

npx claude-flow@alpha sparc run researcher \
  "/prisma-init --topic 'attention mechanisms in vision transformers'"

Deliverables:

• PRISMA protocol document
• Search strategy
• Inclusion/exclusion criteria
• Quality assessment framework

Agent: researcher

---

Pipeline D: Baseline Replication

Objective: Reproduce published baseline with ±1% tolerance

Execution:

claude-code invoke-skill baseline-replication \
  --paper "Attention is All You Need" \
  --dataset ImageNet \
  --tolerance 0.01

Deliverables:

• Baseline implementation (100% test coverage)
• Statistical comparison (±1% tolerance)
• Reproducibility package (Docker)
• Baseline evaluation report

Agents: researcher, data-steward, coder, tester, archivist, evaluator

---

Quality Gate 1 Validation

Objective: GO/NO-GO decision for method development

Execution:

claude-code invoke-skill gate-validation --gate 1

Gate 1 Requirements:

• Literature review complete (≥50 papers)
• Datasheet complete (Form F-P1, ≥90% filled)
• Ethics review APPROVED (data-steward + ethics-agent)
• Baseline replication ±1% tolerance
• Reproducibility package tested (3/3 runs successful)
• Dataset validated (bias audit complete)

Decision:

• APPROVED: Proceed to Phase 2 (Method Development)
• CONDITIONAL: Minor fixes required, proceed with restrictions
• REJECTED: Critical issues, return to Phase 1

Agent: evaluator

---

PHASE 2: DEVELOPMENT (6-12 weeks)

Pipeline D: Method Development (continued)

Objective: Develop novel method with ablation studies

Execution:

claude-code invoke-skill method-development \
  --baseline-checkpoint phase1-foundations/baseline/checkpoint.pth \
  --novel-components "multi-scale-attention,prenorm-residual"

Deliverables:

• Novel method implementation
• Ablation study results (≥5 components)
• Hyperparameter optimization results
• Performance comparison vs. baseline
• Method card (Mitchell et al. 2019 template)

Agents: system-architect, coder, tester, reviewer

---

Pipeline E: Holistic Evaluation

Objective: Comprehensive evaluation across 6+ dimensions

Execution:

claude-code invoke-skill holistic-evaluation \
  --model phase2-development/novel-method/checkpoint.pth \
  --dimensions "accuracy,fairness,robustness,efficiency,interpretability,safety"

Deliverables:

• Holistic evaluation report
• Fairness metrics (demographic parity, equalized odds)
• Robustness analysis (adversarial, OOD)
• Efficiency profiling (latency, memory, energy)
• Interpretability analysis (SHAP, attention viz)
• Safety evaluation (harmful outputs, bias, privacy)

Agents: tester, ethics-agent

---

Pipeline F: Ethics & Safety Review

Objective: Ethics validation for model deployment

Execution:

npx claude-flow@alpha sparc run ethics-agent \
  "/assess-risks --component model --gate 2"

npx claude-flow@alpha sparc run ethics-agent \
  "/safety-eval --model phase2-development/novel-method/checkpoint.pth"

Deliverables:

• Ethics review form (F-F1)
• Risk assessment across 6 domains
• Safety evaluation report
• Fairness audit
• Privacy audit (membership inference)

Agent: ethics-agent

---

Quality Gate 2 Validation

Objective: GO/NO-GO decision for production deployment

Execution:

claude-code invoke-skill gate-validation --gate 2

Gate 2 Requirements:

• Novel method outperforms baseline (statistically significant)
• Ablation studies complete (≥5 components)
• Holistic evaluation complete (6+ dimensions)
• Ethics review APPROVED (ethics-agent)
• Method card complete (≥90% filled)
• Reproducibility tested (3/3 runs successful)

Decision:

• APPROVED: Proceed to Phase 3 (Production)
• CONDITIONAL: Mitigation plan required
• REJECTED: Critical issues, return to Phase 2

Agent: evaluator

---

PHASE 3: PRODUCTION (2-4 weeks)

Pipeline G: Reproducibility & Archival

Objective: Create production-ready reproducibility package

Execution:

# Reproducibility audit
claude-code invoke-skill reproducibility-audit \
  --package phase2-development/novel-method/

# Archival
npx claude-flow@alpha sparc run archivist \
  "/init-model-card --method novel-method --include-metrics"

npx claude-flow@alpha sparc run archivist \
  "Create reproducibility package with Docker, assign DOIs (Zenodo)"

Deliverables:

• Model card (Form F-G2, ≥90% filled)
• Reproducibility package (code + data + environment)
• DOIs assigned (dataset, model, code)
• Registry URLs (HuggingFace, Zenodo)
• Archive (.tar.gz with manifest)

Agent: archivist

---

Pipeline H: Deployment Readiness

Objective: Production deployment validation

Execution:

claude-code invoke-skill deployment-readiness \
  --model phase3-production/final-checkpoint.pth \
  --environment production

Deliverables:

• Deployment checklist
• Infrastructure requirements
• Monitoring plan
• Incident response plan
• Rollback strategy
• Performance benchmarks (production environment)

Agents: tester, archivist

---

Pipeline I: Publication

Objective: Academic paper with reproducibility artifacts

Execution:

claude-code invoke-skill research-publication \
  --results phase1-foundations/ phase2-development/ phase3-production/ \
  --venue "NeurIPS" \
  --artifact-track true

Deliverables:

• Research paper draft
• Reproducibility checklist (NeurIPS, ICML)
• Supplementary materials
• Artifact submission (ACM badges)
• Code release (GitHub)

Agents: researcher, archivist

---

Quality Gate 3 Validation

Objective: Final GO/NO-GO for production deployment and publication

Execution:

claude-code invoke-skill gate-validation --gate 3

Gate 3 Requirements:

• Model card complete (≥90% filled)
• Reproducibility package tested (3/3 runs)
• DOIs assigned (dataset, model, code)
• Code public (GitHub release)
• Ethics review APPROVED (deployment)
• Deployment plan validated
• Publication artifacts ready

Decision:

• APPROVED: Deploy to production, submit publication
• CONDITIONAL: Minor documentation fixes
• REJECTED: Critical reproducibility or ethics issues

Agent: evaluator

---

Deep Research SOP Architecture

3 Phases

Phase 1: FOUNDATIONS (2-4 weeks)
├── Literature Synthesis (Pipeline A)
├── Data & Ethics Foundation (Pipeline B)
├── PRISMA Protocol (Pipeline C, optional)
├── Baseline Replication (Pipeline D)
└── Quality Gate 1 → GO/NO-GO

Phase 2: DEVELOPMENT (6-12 weeks)
├── Method Development (Pipeline D continued)
├── Holistic Evaluation (Pipeline E)
├── Ethics & Safety Review (Pipeline F)
└── Quality Gate 2 → GO/NO-GO

Phase 3: PRODUCTION (2-4 weeks)
├── Reproducibility & Archival (Pipeline G)
├── Deployment Readiness (Pipeline H)
├── Publication (Pipeline I)
└── Quality Gate 3 → GO/NO-GO → DEPLOY

9 Pipelines (A-I)

• Pipeline A: Literature Synthesis
• Pipeline B: Data & Ethics Foundation
• Pipeline C: PRISMA Protocol (systematic reviews)
• Pipeline D: Baseline Replication → Method Development
• Pipeline E: Holistic Evaluation
• Pipeline F: Ethics & Safety Review
• Pipeline G: Reproducibility & Archival
• Pipeline H: Deployment Readiness
• Pipeline I: Publication

3 Quality Gates

• Gate 1: Data & Methods Validation (end of Phase 1)
• Gate 2: Model & Evaluation Validation (end of Phase 2)
• Gate 3: Production & Artifacts Validation (end of Phase 3)

---

Agent Coordination Matrix

Phase	Pipeline	Lead Agent	Supporting Agents
1	A (Literature)	researcher	-
1	B (Data & Ethics)	data-steward	ethics-agent
1	C (PRISMA)	researcher	-
1	D (Baseline)	coder	researcher, tester, archivist
1	Gate 1	evaluator	ALL agents review
2	D (Method Dev)	system-architect	coder, tester, reviewer
2	E (Holistic Eval)	tester	ethics-agent
2	F (Ethics)	ethics-agent	-
2	Gate 2	evaluator	ethics-agent reviews
3	G (Archival)	archivist	-
3	H (Deployment)	tester	archivist
3	I (Publication)	researcher	archivist
3	Gate 3	evaluator	archivist reviews

---

Memory Coordination

Session Persistence

All project state stored in Memory MCP for cross-session coordination:

# Store phase progress
npx claude-flow@alpha memory store \
  --key "sop/project/phase1/status" \
  --value "COMPLETE" \
  --metadata '{"gate1": "APPROVED", "date": "2025-11-01"}'

# Retrieve previous work
npx claude-flow@alpha memory retrieve \
  --key "sop/project/phase1/baseline-results"

# Agent coordination via memory
npx claude-flow@alpha memory store \
  --key "sop/coordination/ethics-agent/status" \
  --value "Awaiting Gate 2 validation" \
  --metadata '{"blocking": ["evaluator"]}'

Cross-Agent Memory Sharing

# data-steward stores bias audit results
npx claude-flow@alpha memory store \
  --key "sop/gate1/bias-audit" \
  --value "$(cat phase1-foundations/bias-audit.json)"

# ethics-agent retrieves for risk assessment
npx claude-flow@alpha memory retrieve \
  --key "sop/gate1/bias-audit"

---

Troubleshooting

Issue: Quality Gate 1 rejected

Symptoms: evaluator returns REJECTED status for Gate 1 Common Causes:

• Baseline replication outside ±1% tolerance
• Incomplete datasheet (<90% filled)
• Ethics review flagged critical data risks
• Reproducibility tests failed

Solutions:

# Check Gate 1 requirements
claude-code invoke-skill gate-validation --gate 1 --verbose

# Re-run baseline replication with debugging
claude-code invoke-skill baseline-replication --debug

# Complete datasheet gaps
npx claude-flow@alpha sparc run data-steward \
  "/init-datasheet --fill-missing-sections"

Issue: Quality Gate 2 rejected

Symptoms: Novel method fails holistic evaluation or ethics review Solutions:

# Review holistic evaluation failures
claude-code invoke-skill holistic-evaluation --dimensions "fairness,safety" --verbose

# Address ethics concerns
npx claude-flow@alpha sparc run ethics-agent \
  "/assess-risks --component model --gate 2 --mitigation-plan"

# Re-run method development with improvements
claude-code invoke-skill method-development --incorporate-feedback

Issue: Quality Gate 3 rejected

Symptoms: Reproducibility package fails or deployment validation issues Solutions:

# Audit reproducibility package
claude-code invoke-skill reproducibility-audit --strict

# Fix deployment issues
claude-code invoke-skill deployment-readiness --fix-issues

# Complete model card
npx claude-flow@alpha sparc run archivist \
  "/init-model-card --complete-missing-sections"

Issue: Phase transitions blocked

Symptoms: Cannot proceed to next phase due to pending validations Solutions:

# Check all gate requirements
npx claude-flow@alpha memory retrieve --key "sop/gates/status"

# Identify blocking agents
npx claude-flow@alpha memory retrieve --key "sop/coordination/*/status"

# Resolve blocking tasks
# (Address specific agent requirements)

---

Integration with Deep Research SOP

Comprehensive Workflow

This orchestrator implements the complete Deep Research SOP as specified in:

• Gap analysis document (identifying missing components)
• 4 P0 commands (/init-datasheet, /prisma-init, /assess-risks, /init-model-card)
• 4 P0 agents (data-steward, ethics-agent, archivist, evaluator)
• 8 GraphViz process diagrams

Quality Assurance

• 3 Quality Gates ensure rigor at each phase transition
• 9 Pipelines provide systematic coverage of research lifecycle
• 4 P0 Agents enforce standards (data, ethics, archival, evaluation)
• Memory MCP enables cross-session coordination and reproducibility

---

Related Skills and Commands

Phase 1 Skills

• literature-synthesis - Systematic literature review
• baseline-replication - Reproduce published baselines

Phase 2 Skills

• method-development - Develop novel methods
• holistic-evaluation - Comprehensive evaluation

Phase 3 Skills

• reproducibility-audit - Audit reproducibility
• deployment-readiness - Production deployment validation
• research-publication - Academic publication

Cross-Phase Skills

• gate-validation - Quality gate validation (all 3 gates)

Related Commands

• /init-datasheet - Create dataset documentation
• /prisma-init - Initialize systematic review
• /assess-risks - Ethics and safety assessment
• /init-model-card - Create model card
• /validate-gate-{1,2,3} - Gate validation

---

References

Deep Research SOP Documentation

• Deep Research SOP Gap Analysis (docs/deep-research-sop-gap-analysis.md)
• 8 GraphViz Process Diagrams (docs/12fa/graphviz/)
• P0 Commands Specification (.claude/commands/research/)
• P0 Agents Specification (agents/research/)

Academic Standards

• Gebru et al. (2021): Datasheets for Datasets
• Mitchell et al. (2019): Model Cards for Model Reporting
• Page et al. (2021): PRISMA 2020 Statement
• ACM Artifact Evaluation Badging
• NeurIPS Reproducibility Checklist

Ethics Frameworks

• IEEE 7010: Well-being Metrics for Ethical AI
• NIST AI Risk Management Framework
• EU AI Act Compliance
• FDA Guidance on AI/ML Medical Devices

---

Appendix

Example Deep Research SOP Timeline

Week 1-2: Phase 1 Start
  - Literature synthesis (50+ papers)
  - Datasheet creation
  - Bias audit

Week 3-4: Phase 1 Complete
  - Baseline replication
  - Ethics review (Gate 1)
  - Gate 1 validation → APPROVED

Week 5-8: Phase 2 Development
  - Novel method implementation
  - Ablation studies
  - Hyperparameter optimization

Week 9-12: Phase 2 Evaluation
  - Holistic evaluation (6 dimensions)
  - Ethics review (Gate 2)
  - Gate 2 validation → APPROVED

Week 13-14: Phase 3 Archival
  - Reproducibility package creation
  - Model card, DOI assignment
  - Registry publishing

Week 15-16: Phase 3 Deployment & Publication
  - Deployment readiness validation
  - Paper writing
  - Gate 3 validation → APPROVED → DEPLOY

Total: 16 weeks (4 months) for complete research lifecycle

Quality Gate Decision Matrix

Gate	APPROVED	CONDITIONAL	REJECTED
Gate 1	All requirements met, proceed to Phase 2	Minor datasheet gaps, proceed with restrictions	Baseline >±1% or critical ethics issues
Gate 2	All requirements met, proceed to Phase 3	Mitigation plan for fairness/robustness gaps	Performance regression or critical safety risks
Gate 3	All requirements met, DEPLOY to production	Minor documentation fixes required	Reproducibility failures or ethics violations