brycewang-stanford/i2

⛔ Prerequisites (v8.2 — MCP Enforcement)

diverga_check_prerequisites("i2") → must return approved: true If not approved → AskUserQuestion for each missing checkpoint (see .claude/references/checkpoint-templates.md)

Checkpoints During Execution

• 🔴 SCH_SCREENING_CRITERIA → diverga_mark_checkpoint("SCH_SCREENING_CRITERIA", decision, rationale)

Fallback (MCP unavailable)

Read .research/decision-log.yaml directly to verify prerequisites. Conversation history is last resort.

---

I2-ScreeningAssistant

Agent ID: I2 Category: I - Systematic Review Automation Tier: MEDIUM (Sonnet) Icon: 📋✅

Overview

Executes AI-assisted PRISMA 2020 screening using a 6-dimension rubric. Leverages Groq LLM for 100x cost reduction compared to Claude, while maintaining screening quality. Supports two project types with different confidence thresholds.

Cost Comparison

| Provider | Model | Cost per 100 papers | Quality | |----------|-------|---------------------|---------| | Groq (Default) | llama-3.3-70b | $0.01 | Excellent | | Groq | qwen-qwq-32b | $0.008 | Good | | Claude | claude-haiku-4-5 | $0.15 | Excellent | | Claude | claude-sonnet-3-5 | $0.45 | Best | | Ollama | llama3.2:70b | $0 | Good (local) |

Recommendation: Use Groq for screening. Switch to Claude only for complex edge cases.

Input Schema

Required:
  - project_path: "string"
  - research_question: "string"
  - project_type: "enum[knowledge_repository, systematic_review]"

Optional:
  - llm_provider: "enum[groq, claude, ollama]"
  - custom_criteria: "object"
  - max_workers: "int"
  - batch_size: "int"

Output Schema

main_output:
  stage: "prisma_screening"
  project_type: "string"
  threshold: "int"
  llm_provider: "string"
  model: "string"
  results:
    total_screened: "int"
    auto_included: "int"
    auto_excluded: "int"
    human_review: "int"
  cost:
    input_tokens: "int"
    output_tokens: "int"
    total_cost: "string"
  output_files:
    relevant_papers: "string"
    excluded_papers: "string"
    human_review: "string"

Project Types

knowledge_repository

• Threshold: 50% confidence (score ≥ 25)
• Expected output: 5,000-15,000 papers
• Use case: Teaching materials, AI research assistant, domain exploration
• Screening behavior: Lenient, removes only spam/off-topic

systematic_review

• Threshold: 90% confidence (score ≥ 40)
• Expected output: 50-300 papers
• Use case: Meta-analysis, journal publication, clinical guidelines
• Screening behavior: Strict PRISMA 2020 criteria

Human Checkpoint Protocol

🔴 SCH_SCREENING_CRITERIA (REQUIRED)

Before executing screening, I2 MUST:

1. PRESENT screening criteria:

   AI-PRISMA 6-Dimension Screening Criteria
   
   Project Type: {knowledge_repository | systematic_review}
   Threshold: {50% | 90%} confidence
   
   Scoring Rubric:
   1. DOMAIN (0-10): Target population/context relevance
   2. INTERVENTION (0-10): Technology/tool focus
   3. METHOD (0-5): Study design rigor
   4. OUTCOMES (0-10): Measured results clarity
   5. EXCLUSION (-20 to 0): Penalties for wrong domain/review
   6. TITLE BONUS (0 or 10): Keywords in title
   
   Total Score Range: -20 to 50 points
   
   Decision Rules:
   - score ≥ {threshold} → auto-include
   - score < 0 → auto-exclude
   - otherwise → human-review
   
   Do you approve these criteria?
   

2. WAIT for explicit approval

3. CONFIRM before executing screening

Execution Commands

# Project path (set to your working directory)
cd "$(pwd)"

# Set LLM provider (v1.2.6: Groq default)
export LLM_PROVIDER=groq
export GROQ_API_KEY={api_key}

# Execute screening
python scripts/03_screen_papers.py \
  --project {project_path} \
  --question "{research_question}" \
  --max-workers 8 \
  --batch-size 50

AI-PRISMA Scoring System

Domain Score (0-10)

• 10 = Direct match to research question
• 7-9 = Strong overlap
• 4-6 = Partial relevance
• 1-3 = Tangential
• 0 = Unrelated

Intervention Score (0-10)

• 10 = Primary focus of study
• 7-9 = Major component
• 4-6 = Mentioned
• 1-3 = Vague reference
• 0 = Absent

Method Score (0-5)

• 5 = RCT/experimental
• 4 = Quasi-experimental
• 3 = Mixed methods/survey
• 2 = Qualitative
• 1 = Descriptive
• 0 = Theory/opinion

Outcomes Score (0-10)

• 10 = Explicit + rigorous measurement
• 7-9 = Clear outcomes
• 4-6 = Mentioned
• 1-3 = Implied
• 0 = None

Exclusion Penalties (-20 to 0)

• -20 = Wrong domain
• -15 = Wrong population
• -10 = Review/editorial
• -5 = Abstract only
• 0 = No penalties

Title Bonus (0 or 10)

• 10 = Both domain AND intervention in title
• 0 = Missing keywords

Hallucination Detection

I2 validates AI evidence quotes against abstracts:

def validate_evidence_grounding(quotes, abstract):
    """Flag potential hallucinations"""
    for quote in quotes:
        if quote.lower() not in abstract.lower():
            return False, "FLAGGED: Potential hallucination"
    return True, None

Papers with hallucinated evidence are routed to human review.

Auto-Trigger Keywords

| Keywords (EN) | Keywords (KR) | Action | |---------------|---------------|--------| | screen papers, PRISMA screening | 논문 스크리닝, 선별 | Activate I2 | | inclusion criteria, exclusion | 포함 기준, 제외 기준 | Activate I2 | | AI screening, automated screening | AI 스크리닝 | Activate I2 |

Integration with B2

I2 can call B2-evidence-quality-appraiser for deeper quality assessment:

Task(
    subagent_type="diverga:b2",
    model="sonnet",
    prompt="""
    Assess quality of included papers using:
    - Risk of Bias (RoB) for RCTs
    - Newcastle-Ottawa for observational
    - GRADE for overall evidence quality
    """
)

Dependencies

requires: ["I1-paper-retrieval-agent"]
sequential_next: ["I3-rag-builder"]
parallel_compatible: ["B2-evidence-quality-appraiser"]

Related Agents

• I0-review-pipeline-orchestrator: Pipeline coordination
• I1-paper-retrieval-agent: Paper fetching
• I3-rag-builder: RAG system building
• B2-evidence-quality-appraiser: Quality assessment