seldon-research

name	seldon-research
description	Deep research using ChatGPT + NotebookLM + domain knowledge — cross-model validation for high-quality answers and continuous improvement

Seldon Research — Cross-Model Deep Research

Combines ChatGPT (second opinion), NotebookLM (institutional memory), and GA domain knowledge to produce high-quality, validated answers. Every research interaction compounds back into the governance framework.

Usage

/seldon research [question] — full research pipeline /seldon research quick [question] — fast answer without cross-model validation

Research Pipeline

1. Classify question (governance/experiential/domain)
       ↓
2. Check NotebookLM for existing knowledge
       ↓
3. Query GA domain skills if music-related
       ↓
4. Cross-validate with ChatGPT (GPT-4o-mini)
       ↓
5. Assess confidence via tetravalent logic
       ↓
6. Deliver answer with full agentic trace
       ↓
7. Compound: log interaction, update evolution, detect patterns

Step 1: Classify

Determine knowledge layer:

• Governance: constitutional articles, policies, compliance → use Demerzel artifacts directly
• Experiential: PDCA outcomes, past decisions, patterns → check evolution log + NotebookLM
• Domain (music): theory, chords, scales → GA skills + NotebookLM Compound the Compounding

Step 2: NotebookLM Research

Query the most relevant notebook:

mcp__notebooklm__ask_question({
  question: "[processed question]",
  notebook_id: "[best match]"  // compound-the-compounding, probabilistic-grammars, etc.
})

If the notebook has relevant context, incorporate it into the answer with citations.

Step 3: GA Domain Skills

For music theory questions, invoke GA skills conceptually:

• ScaleInfoSkill → scale notes, patterns, formulas
• ChordExplanationSkill → chord construction, qualities
• IntervalInfoSkill → interval identification
• ModeExplorationSkill → modal theory
• ProgressionSuggestionSkill → chord progressions
• HarmonicAnalysisSkill → harmonic function analysis

These produce deterministic answers (confidence 1.0).

Step 4: ChatGPT Cross-Validation

For non-trivial questions, ask GPT-4o-mini to validate:

mcp__openai-chat__openai_chat({
  model: "gpt-4o-mini",
  messages: [
    { role: "system", content: "You are a music theory expert. Verify this answer is accurate. Be concise." },
    { role: "user", content: "[answer to validate]" }
  ]
})

• If ChatGPT agrees → confidence increases
• If ChatGPT disagrees → belief state becomes Contradictory (C) → investigate
• If ChatGPT adds useful context → incorporate with attribution

Step 5: Tetravalent Assessment

Rate the answer:

• T (True): Domain computation + ChatGPT agreement + NotebookLM confirmation
• U (Unknown): Partial evidence, needs more research
• C (Contradictory): Sources disagree → flag for human review
• F (False): Answer debunked → do not deliver

Step 6: Deliver with Trace

Include full agentic trace showing which sources were used, which models validated, and confidence level.

Step 7: Compound

Every research interaction feeds back:

• Citation tracking: Which governance artifacts were used? Update evolution log.
• Pattern detection: Is this topic asked frequently? Candidate for dedicated resource.
• Knowledge gap detection: Did we fail to answer? Gap in domain coverage → create issue.
• Improvement proposals: Could the answer be better? Log as Kaizen opportunity.
• Cross-repo learning: Did this reveal something about GA/TARS/IX? Share via Galactic Protocol.

Compounding Rules

1. Every unanswered question = governance gap → create issue in relevant repo
2. Every answered question = citation → update evolution log for Seldon + Streeling
3. Every ChatGPT disagreement = investigation → log as Contradictory, escalate if serious
4. Every 3+ questions on same topic = promotion candidate → propose dedicated guide
5. Every failed answer = experiential learning → package as knowledge state for Seldon

Cost Control (ChatGPT)

• Use GPT-4o-mini only (~$0.15/M input)
• Only cross-validate non-trivial questions (skip for basic chord/scale lookups)
• Max 1 ChatGPT call per question
• Never send user PII or credentials

Source

personas/seldon.persona.yaml, policies/streeling-policy.yaml, policies/scientific-objectivity-policy.yaml (confidence calibration)