# Structured Extraction Template
# Machine-readable extraction of claims, methods, findings from research papers

---
template_id: extraction
version: 1.0.0
reasoning_required: true
framework: research-complete
format: yaml
---

# USAGE NOTES:
# This template extracts structured data from papers for machine processing.
# Complete REASONING section in markdown comment first, then populate YAML fields.
# Use this for: citation networks, claim verification, method comparison, finding synthesis.

# REASONING (Complete before extraction):
#
# 1. **Extraction Scope**: What information do we need from this paper?
#    > [Define which claims, methods, findings are relevant to AIWG - not everything]
#
#    EXAMPLE:
#    For REF-018 ReAct, extract: TAO loop structure, performance metrics on tool-use tasks,
#    hallucination reduction data, reasoning trace format. Skip: Benchmark-specific details
#    not applicable to SDLC workflows.
#
# 2. **Claim Classification**: How do we categorize extracted claims?
#    > [Decide taxonomy: empirical vs theoretical, quantitative vs qualitative, causal vs correlational]
#
#    EXAMPLE:
#    - Empirical claims: "ReAct reduces hallucinations to 0%" (measured)
#    - Theoretical claims: "Reasoning traces enable better oversight" (proposed)
#    - Causal claims: "TAO loop CAUSES performance improvement" (proven causation)
#    - Correlational claims: "TAO loop ASSOCIATED with fewer errors" (correlation observed)
#
# 3. **Evidence Strength Assessment**: How do we rate evidence quality?
#    > [Use GRADE levels: HIGH/MODERATE/LOW/VERY LOW based on methodology]
#
#    EXAMPLE:
#    HIGH: Controlled experiment, multiple baselines, reproducible = "34% improvement"
#    MODERATE: Single comparison, limited scope = "Improved performance observed"
#    LOW: Case study, no controls = "Practitioner reports benefits"
#
# 4. **Method Abstraction**: What level of detail for methods?
#    > [Balance: enough detail to evaluate applicability, not so much it's unusable]
#
#    EXAMPLE:
#    Extract: Core algorithm (TAO loop), key parameters (max iterations), evaluation approach
#    Skip: Benchmark-specific implementation details, dataset preprocessing minutiae
#
# 5. **Applicability Mapping**: How does this apply to AIWG?
#    > [For each extraction, note AIWG component(s) affected]
#
#    EXAMPLE:
#    TAO loop → All tool-using agents (@.claude/rules/tao-loop.md)
#    Thought types → Agent reasoning (@.claude/rules/thought-protocol.md)
#    Hallucination reduction → Citation verification (@.claude/rules/citation-policy.md)

metadata:
  extraction_id: "extraction-REF-XXX"
  paper_ref: "REF-XXX"
  paper_title: "Full Paper Title"
  authors:
    - "Author 1"
    - "Author 2"
  year: YYYY
  extraction_date: "YYYY-MM-DD"
  extractor: "Agent or Human Name"
  extraction_confidence: 0.95  # 0-1 scale

  # EXAMPLE:
  # extraction_id: "extraction-REF-018"
  # paper_ref: "REF-018"
  # paper_title: "ReAct: Synergizing Reasoning and Acting in Language Models"
  # authors:
  #   - "Yao, S."
  #   - "Zhao, J."
  #   - "Yu, D."
  # year: 2022
  # extraction_date: "2026-02-03"
  # extractor: "Discovery Agent"
  # extraction_confidence: 0.95

claims:
  # List of specific claims made in the paper with evidence
  # Each claim is independently verifiable

  - claim_id: "claim-001"
    claim_text: "[Exact claim or close paraphrase]"
    claim_type: "empirical"  # empirical | theoretical | methodological
    evidence_type: "quantitative"  # quantitative | qualitative | mixed
    causality: "causal"  # causal | correlational | descriptive

    # EXAMPLE:
    # claim_id: "claim-001"
    # claim_text: "ReAct improves success rate by 34% on HotpotQA compared to Act-only baseline"
    # claim_type: "empirical"
    # evidence_type: "quantitative"
    # causality: "causal"

    evidence:
      data_points:
        - metric: "[Metric name]"
          baseline: "[Baseline value]"
          result: "[Result value]"
          improvement: "[% or absolute improvement]"
          # EXAMPLE:
          # metric: "HotpotQA Success Rate"
          # baseline: "49%"
          # result: "66%"
          # improvement: "+34% relative"

      source_location:
        page: "X"
        section: "Section Name"
        figure_table: "Figure/Table Y"
        # EXAMPLE:
        # page: "4"
        # section: "4.1 Question Answering Results"
        # figure_table: "Table 1"

      quote: "[Exact quote from paper if available]"
      # EXAMPLE:
      # quote: "ReAct achieves 66% success rate on HotpotQA, compared to 49% for Act-only, representing a 34% relative improvement"

    grade_quality: "HIGH"  # HIGH | MODERATE | LOW | VERY_LOW
    quality_rationale: "[Why this quality rating?]"
    # EXAMPLE:
    # grade_quality: "HIGH"
    # quality_rationale: "Controlled experiment, clear baselines, multiple task evaluation, reproducible methodology"

    applicability_to_aiwg:
      relevance: "HIGH"  # HIGH | MODERATE | LOW | NONE
      affected_components:
        - "@.claude/rules/tao-loop.md"
        - "@agentic/code/frameworks/sdlc-complete/agents/*.md"
      implementation_notes: "[How this applies to AIWG]"
      # EXAMPLE:
      # relevance: "HIGH"
      # affected_components:
      #   - "@.claude/rules/tao-loop.md"
      #   - "@agentic/code/frameworks/sdlc-complete/agents/requirements-analyst.md"
      # implementation_notes: "Standardize TAO loop across all tool-using agents to achieve similar performance gains"

    limitations:
      - "[Limitation 1 of this claim]"
      - "[Limitation 2 of this claim]"
      # EXAMPLE:
      # - "Tested only on QA tasks, not full SDLC workflows"
      # - "Single-agent context, multi-agent coordination not evaluated"

  # ANTI-PATTERN EXAMPLE:
  # - claim_id: "claim-bad"
  #   claim_text: "The method works well"  # Too vague
  #   claim_type: "empirical"
  #   evidence_type: "qualitative"
  #   causality: "descriptive"
  #   evidence:
  #     quote: "It improved performance"  # No specifics
  #   grade_quality: "LOW"

  # BETTER:
  # - claim_id: "claim-good"
  #   claim_text: "ReAct reduces hallucination rate from 56% to 0% on FEVER fact verification"
  #   claim_type: "empirical"
  #   evidence_type: "quantitative"
  #   causality: "causal"
  #   evidence:
  #     data_points:
  #       - metric: "Hallucination Rate"
  #         baseline: "56%"
  #         result: "0%"
  #         improvement: "100% reduction"
  #     source_location:
  #       page: "6"
  #       figure_table: "Figure 3"

methods:
  # Methodologies, algorithms, techniques introduced or used

  - method_id: "method-001"
    method_name: "[Short descriptive name]"
    method_type: "algorithm"  # algorithm | framework | evaluation_protocol | experimental_design

    # EXAMPLE:
    # method_id: "method-001"
    # method_name: "ReAct Loop (Thought→Action→Observation)"
    # method_type: "algorithm"

    description: "[Detailed description of the method]"
    # EXAMPLE:
    # description: "Iterative loop interleaving reasoning traces with tool actions and observations. Each iteration: 1) THOUGHT - reasoning about current state, 2) ACTION - tool invocation, 3) OBSERVATION - result capture."

    pseudocode: |
      # Optional: pseudocode representation
      # EXAMPLE:
      # while not task_complete and iterations < max:
      #   thought = generate_reasoning(state)
      #   action = select_tool_and_params(thought)
      #   observation = execute_tool(action)
      #   state = update_state(observation)

    key_parameters:
      - param: "[Parameter name]"
        value: "[Value or range]"
        description: "[What this parameter controls]"
        # EXAMPLE:
        # param: "max_iterations"
        # value: "5-10"
        # description: "Maximum TAO loop iterations before termination"

      - param: "[Parameter name]"
        value: "[Value or range]"
        description: "[What this parameter controls]"
        # EXAMPLE:
        # param: "temperature"
        # value: "0.7"
        # description: "LLM sampling temperature for thought generation"

    evaluation:
      benchmarks:
        - name: "[Benchmark name]"
          result: "[Result on this benchmark]"
          # EXAMPLE:
          # name: "HotpotQA"
          # result: "66% success rate"

      baselines:
        - name: "[Baseline name]"
          result: "[Baseline result]"
          comparison: "[How method compares]"
          # EXAMPLE:
          # name: "Act-only (no reasoning)"
          # result: "49% success rate"
          # comparison: "+34% improvement with ReAct"

    reproducibility:
      code_available: true  # true | false
      code_url: "https://github.com/..."
      data_available: true  # true | false
      data_url: "https://..."
      # EXAMPLE:
      # code_available: true
      # code_url: "https://github.com/ysymyth/ReAct"
      # data_available: true
      # data_url: "https://hotpotqa.github.io/"

    applicability_to_aiwg:
      can_implement: true  # true | false | partial
      implementation_complexity: "moderate"  # low | moderate | high
      dependencies:
        - "[Dependency 1]"
        - "[Dependency 2]"
      # EXAMPLE:
      # can_implement: true
      # implementation_complexity: "moderate"
      # dependencies:
      #   - "LLM API (GPT-4, Claude, etc.)"
      #   - "Tool execution environment (Bash, Read, Write, etc.)"

      implementation_status:
        - component: "[AIWG component]"
          status: "implemented"  # planned | in_progress | implemented | not_applicable
          reference: "@path/to/implementation"
          # EXAMPLE:
          # component: "TAO Loop Standardization"
          # status: "implemented"
          # reference: "@.claude/rules/tao-loop.md"

findings:
  # Key findings, insights, discoveries from the research

  - finding_id: "finding-001"
    finding_text: "[Clear statement of the finding]"
    finding_type: "performance"  # performance | insight | limitation | recommendation

    # EXAMPLE:
    # finding_id: "finding-001"
    # finding_text: "Tool grounding (external observations) reduces hallucinations to near-zero"
    # finding_type: "insight"

    supporting_evidence:
      - claim_ref: "claim-001"  # Reference to claim ID above
      - claim_ref: "claim-002"
      # EXAMPLE:
      # - claim_ref: "claim-002"  # (hypothetical) "0% hallucinations with tool use"

    significance: "HIGH"  # HIGH | MODERATE | LOW
    significance_rationale: "[Why this finding matters]"
    # EXAMPLE:
    # significance: "HIGH"
    # significance_rationale: "Directly addresses critical failure mode in LLM systems. Provides actionable pattern for reducing fabricated information."

    implications:
      - domain: "[Domain this affects]"
        implication: "[What this means for that domain]"
        # EXAMPLE:
        # domain: "Agent Reliability"
        # implication: "Agents must ground claims in tool observations (Read, Grep results) rather than generating from parametric knowledge alone"

      - domain: "[Domain this affects]"
        implication: "[What this means for that domain]"
        # EXAMPLE:
        # domain: "Citation Verification"
        # implication: "Before citing sources, agents must use Read tool to verify file exists and extract exact quote"

    limitations:
      - "[Limitation 1]"
      - "[Limitation 2]"
      # EXAMPLE:
      # - "Finding based on QA tasks; applicability to code generation unknown"
      # - "Tool reliability assumed; unreliable tools may introduce new errors"

    future_work:
      - "[Research gap 1]"
      - "[Research gap 2]"
      # EXAMPLE:
      # - "Evaluate tool grounding in multi-agent coordination scenarios"
      # - "Develop tool reliability metrics and selection strategies"

relationships:
  # Connections to other papers in the corpus

  builds_on:
    - paper_ref: "REF-XXX"
      relationship: "[How this builds on that paper]"
      # EXAMPLE:
      # paper_ref: "REF-016"
      # relationship: "Extends Chain-of-Thought by adding action execution and observation phases"

  extends:
    - paper_ref: "REF-XXX"
      relationship: "[How this extends that paper]"
      # EXAMPLE:
      # paper_ref: "REF-019"
      # relationship: "Extends Toolformer by adding explicit reasoning traces before tool use"

  contradicts:
    - paper_ref: "REF-XXX"
      relationship: "[How this contradicts that paper]"
      contradiction_type: "methodology"  # methodology | findings | interpretation
      # EXAMPLE:
      # paper_ref: "REF-XXX"
      # relationship: "Contradicts assumption that more tool use always improves performance; shows reasoning-first is critical"
      # contradiction_type: "findings"

  cited_by:
    - paper_ref: "REF-XXX"
      relationship: "[How that paper uses this work]"
      # EXAMPLE:
      # paper_ref: "REF-022"
      # relationship: "AutoGen adopts ReAct patterns for agent communication"

synthesis:
  # High-level synthesis across claims, methods, findings

  core_contribution: "[The single most important contribution of this paper]"
  # EXAMPLE:
  # core_contribution: "Demonstrates that interleaving reasoning with actions significantly improves LLM task performance and enables tool grounding that eliminates hallucinations"

  practical_takeaways:
    - "[Actionable takeaway 1]"
    - "[Actionable takeaway 2]"
    - "[Actionable takeaway 3]"
    # EXAMPLE:
    # - "Implement TAO loop structure in all agents that use tools"
    # - "Track thought types (goal, progress, extraction, reasoning, exception, synthesis)"
    # - "Require agents to ground claims in tool observations before stating facts"

  open_questions:
    - "[Unanswered question 1]"
    - "[Unanswered question 2]"
    # EXAMPLE:
    # - "How does TAO loop scale to 10+ iteration sessions (Ralph loops)?"
    # - "Can reasoning quality be measured automatically?"
    # - "How do multiple agents coordinate with ReAct patterns?"

  confidence_assessment:
    overall_confidence: 0.90  # 0-1 scale
    confidence_factors:
      methodology_rigor: 0.95
      evidence_strength: 0.90
      generalizability: 0.85
    confidence_notes: "[Why this confidence level?]"
    # EXAMPLE:
    # overall_confidence: 0.90
    # confidence_factors:
    #   methodology_rigor: 0.95  # Excellent experimental design
    #   evidence_strength: 0.90  # Strong quantitative results
    #   generalizability: 0.85  # QA tasks, not full SDLC workflows
    # confidence_notes: "High confidence in findings for tool-use tasks; moderate confidence for complex SDLC workflows"

references:
  # Links to related artifacts
  literature_note: "@.aiwg/research/findings/REF-XXX.md"
  summary: "@.aiwg/research/summaries/REF-XXX-summary.md"
  source_pdf: "@.aiwg/research/sources/REF-XXX.pdf"
  provenance_record: "@.aiwg/research/provenance/records/REF-XXX.prov.yaml"

  aiwg_implementations:
    - "@.claude/rules/tao-loop.md"
    - "@.claude/rules/thought-protocol.md"
    - "@agentic/code/frameworks/sdlc-complete/agents/*.md"

  # EXAMPLE:
  # literature_note: "@.aiwg/research/findings/REF-018-react.md"
  # summary: "@.aiwg/research/summaries/REF-018-summary.md"
  # source_pdf: "@.aiwg/research/sources/yao-2022-react.pdf"
  # provenance_record: "@.aiwg/research/provenance/records/REF-018.prov.yaml"
  # aiwg_implementations:
  #   - "@.claude/rules/tao-loop.md"
  #   - "@.claude/rules/thought-protocol.md"

metadata_schema_version: "1.0.0"
extraction_schema: "@agentic/code/frameworks/research-complete/schemas/extraction-schema.yaml"

# VALIDATION CHECKLIST (verify before finalizing):
# [ ] All claims have evidence with source location
# [ ] All claims have GRADE quality assessment
# [ ] All methods have applicability to AIWG
# [ ] All findings reference supporting claims
# [ ] Relationships to other papers documented
# [ ] Confidence assessment completed with rationale
# [ ] References to AIWG implementations included
# [ ] No vague claims ("improves performance" → specify metric and magnitude)
# [ ] No unsupported quality ratings (justify HIGH/MODERATE/LOW)

# ANTI-PATTERNS TO AVOID:
# ❌ Extracting every claim (focus on AIWG-relevant only)
# ❌ Vague evidence ("page 5" without specific quote or data)
# ❌ Missing quality assessment (every claim needs GRADE level)
# ❌ Copy-pasting abstract (synthesize, don't duplicate)
# ❌ No applicability notes (always map to AIWG components)