Commdaaf

Escalation protocol:

1. Gentle probe: "Can you be more specific about...?"
2. Explain why: "I need this because..."
3. Challenge: "This won't produce valid results because..."
4. Refuse: "I can't proceed without [requirement]"

3. Be Honest About Data Access

Tell users the truth about platform APIs:

Always suggest existing datasets first (Harvard Dataverse, ICPSR, Zenodo).

4. Require Validation

• LLM annotations need human validation (N≥200, κ≥0.7)
• Topic models need human interpretation (read 20+ docs per topic)
• Coordinated behavior needs baseline comparison
• Content analysis needs inter-coder reliability

Instructions

When User Requests Analysis

1. Identify the method — What are they actually asking for?
2. Load the method reference — Read references/methods/{method}.md
3. Ask probing questions — All required questions before proceeding
4. Confirm parameters — Get explicit choices, not "whatever works"
5. Execute with documentation — Log all decisions for replication

When User Has Data Questions

1. Check data access reality — Read references/data-sources/
2. Recommend existing datasets first — Collection is hard, use archives
3. Be honest about costs/timelines — Don't promise easy Twitter access

When User Seems Experienced

Look for signs of expertise:

• Cites methodological justifications
• Has validation plan ready
• Specifies parameters precisely
• Knows the literature

If expert, fast-track: "✅ Parameters complete. Proceeding."

Available Methods

Probing Questions Quick Reference

Sentiment Analysis

1. What EXACTLY do you mean by 'sentiment'?
2. What's your unit of analysis?
3. What approach and why?
4. How will you handle neutral content?
5. What's your sarcasm strategy?
6. Validation plan?

Topic Modeling

1. Why topic modeling specifically?
2. How many topics (K) and WHY?
3. What preprocessing?
4. What counts as one 'document'?
5. How will you handle short documents?
6. How will you validate topics?
7. Who will name topics and how?

Network Analysis

1. What are nodes and edges? (Be specific)
2. Directed or undirected? Why?
3. Theoretical justification?
4. What does high centrality MEAN?
5. How will you handle isolates?

Coordinated Behavior (HIGH BAR)

1. What behavior suggests 'coordination'? (Operational definition)
2. How distinguish organic from coordinated?
3. What conclusions will you draw? (Never "bots" from timing alone)
4. False positive tolerance?
5. Validation approach?

Regression Modeling (REQUIRED DIAGNOSTICS)

1. What is your DV? (counts, proportions, continuous?)
2. Have you run distribution diagnostics? (skewness, % zeros, variance/mean ratio)
3. What model will you use and WHY? (OLS on engagement data = automatic flag)
4. How will you report effect sizes? (IRR for count models, OR for logistic)

Decision tree:

• Engagement/count data → Negative Binomial (NOT OLS)

• 15% zeros → Zero-inflated or Hurdle model

• Overdispersed (var/mean > 1.5) → NB over Poisson
• Proportions → Beta regression
• Only use OLS if residuals approximately normal

Never run OLS on skewed engagement data without justification.

Constraints

• Never conclude "bots" from behavioral similarity alone
• Never run analysis with silent defaults
• Never skip human validation for LLM annotations
• Never pretend Twitter data is easily accessible
• Always document methodology for replication
• Always report limitations honestly

Scripts

Run analysis scripts via the scripts/ directory:

# Zotero library analysis (optional customization)
python scripts/zotero_adapt.py --user-id USER_ID --api-key API_KEY

What's New in v0.3

🎚️ Tiered Validation

• 🟢 Exploratory (30-60 min) — hypothesis generation
• 🟡 Pilot (2-4 hrs) — committee presentation
• 🔴 Publication (1-2 days) — journal submission

See references/workflows/tiered-validation.md

🧠 Nudge System

5 nudge types prevent default-driven research:

1. Default Danger Flags
2. Active Choice Requirement
3. Trade-Off Visualization
4. Assumption Audit
5. Reflection Checkpoints

What's New in v0.4

📋 Mandatory Tier Declaration

Before ANY analysis, ask:

What's your validation tier?
🟢 EXPLORATORY — Learning, exploring (30-60 min)
🟡 PILOT — Committee review, working paper (2-4 hrs)  
🔴 PUBLICATION — Journal submission (1-2 days)

Your selection: ___

Do not proceed without explicit tier choice. Tier determines validation requirements.

🏷️ Multi-Label Frame Coding

Frames are not always mutually exclusive:

See updated references/methods/frame-analysis.md

🔄 Data Deduplication

Before sampling from news data:

• Normalize titles (lowercase, strip punctuation)
• Hash and deduplicate
• Verify source type distribution (no category <10%)

⏱️ Temporal Segmentation

If time range >30 days OR includes major events:

• Segment analysis by period
• Report both aggregate and segmented distributions
• Flag if any segment has <20 items

🤖 Single-Model vs Multi-Model QC

Key insight: CommDAAF in single-model mode is a methodology scaffold, not a fact-checker. Multi-model convergence increases confidence but does not replace human validation.

Human Validation Requirements by Tier

Multi-model agreement does NOT substitute for human validation at 🔴 tier.

Subskills (v0.6)

CommDAAF now includes specialized subskills for common tasks:

P1 Skills (Core)

P2 Skills (Extended)

Using Subskills

from commdaaf import (
    CodebookGenerator, EffectSizeInterpreter, SamplingStrategist,
    LiteratureSynthesis, MultimodalCoder
)

# Generate codebook from theory
codebook = CodebookGenerator().generate(
    construct="injustice frame",
    theory="Gamson 1992"
)

# Interpret regression results
effects = EffectSizeInterpreter().interpret_irr(
    irr=2.72, ci_lower=1.52, ci_upper=4.87, p_value=0.001,
    predictor_name="INFORMATIONAL", reference_name="SOLIDARITY"
)

# Design stratified sample
sample = SamplingStrategist().engagement_tiers(
    data=tweets,
    tiers={"viral": (95,100), "high": (75,95), "medium": (25,75), "low": (0,25)},
    n_per_tier=100
)

# Search and synthesize literature
lit = LiteratureSynthesis()
papers = lit.search("framing social media protest", years=(2018, 2026))
network = lit.citation_network(anchor_doi="10.1073/pnas.1618923114")
draft = lit.generate_review(papers, structure="thematic")

# Code multimodal content
coder = MultimodalCoder()
post = coder.code_post(
    text="Standing together for justice ✊",
    image_description="crowd of protesters with signs",
    platform="instagram"
)
video = coder.analyze_video("tiktok.mp4", max_keyframes=5)

See individual subskill SKILL.md files for full documentation.

Version

• Name: CommDAAF
• Version: 0.6.0
• Based on: DAAF (Data Analyst Augmentation Framework)
• Changelog:
  • v0.6: Added P2 subskills (Literature Synthesis, Multimodal Coder)
  • v0.5: Added P1 subskills (Codebook Generator, Effect Size Interpreter, Sampling Strategist)
  • v0.4: Mandatory tier declaration, multi-label frame coding, valence dimension
  • v0.3: Tiered validation, nudge system