Self Improving Investigation

Vasana

A vasana is a pattern that persists across unrelated contexts. If during this task you notice such a pattern emerging, it may be worth capturing. This skill works best alongside the vasana skill and vasana hook from the Vasana System plugin.

Modify freely. Keep this section intact.

Self-Improving Investigation Methodology

Seed Question: Am I investigating, or just confirming what I already believe?

Status: DESCRIPTIVE, NOT PRESCRIPTIVE

Version: 1.2 (2026-03-09) - Added source diversity framework integration, worker source assignment, dialectic source category checking

This methodology emerged from actual use, not theory. It documents what worked during real investigation, with notes on what could be improved. It is open to further improvement through iteration.

Treat this as a starting point, not a rulebook. If something doesn't work for your task, adapt it and document what you learned.

Recent Improvements (v1.2)

• Source diversity in worker assignment - Assign workers sources from DIFFERENT categories to prevent perspective monoculture
• Dialectic source category check - Verify thesis and antithesis come from different source categories
• Integration with SOURCE_DIVERSITY_FRAMEWORK.md - Full source taxonomy for multi-perspective research

Previous Improvements (v1.1)

• Parallel loops - Run independent research dimensions simultaneously to prevent bias transfer
• Multiple antitheses - Seek 4+ antitheses, not just one; each challenges different assumption
• Conceptual distinction emergence - Watch for vague terms that need splitting into distinct concepts

Core Architecture

The Problem Being Solved

Standard research suffers from:

1. Confirmation bias - Finding evidence for pre-existing beliefs
2. Premature exit - Stopping when "good enough" instead of when criteria pass
3. Context pollution - Previous conclusions influencing new research
4. Single-perspective analysis - Missing counter-arguments and alternatives

The Solution: Three-Layer Architecture

┌─────────────────────────────────────────────────────────┐
│ LAYER 1: ORCHESTRATION                                  │
│ (Main session - sees everything, intervenes minimally)  │
├─────────────────────────────────────────────────────────┤
│ LAYER 2: BLIND WORKERS                                  │
│ (Fresh context per task - no awareness of orchestration)│
├─────────────────────────────────────────────────────────┤
│ LAYER 3: DIALECTIC SYNTHESIS                           │
│ (Thesis → Antithesis → Synthesis for each major claim) │
└─────────────────────────────────────────────────────────┘

Layer 1: Orchestration

Role: Observe, judge, minimally intervene.

Orchestrator Responsibilities

1. Define completion criteria BEFORE starting
2. Spawn blind workers with task-only prompts
3. Read worker output (one-way data flow)
4. Judge whether intervention needed
5. Log orchestrator decisions for methodology improvement
6. Synthesize findings using dialectic method

Minimal Intervention Principle

Ideal: Prompt stays EXACTLY the same across iterations. Reality: Sometimes minimal changes needed - but RESIST the urge.

Layer 2: Blind Workers

Critical: Workers must be BLIND to orchestration.

What Workers DON'T Know

• That they're part of a loop
• That they're called "workers"
• What iteration this is
• That logs exist
• That orchestration is happening

Worker Prompt Pattern

[Task description only]

Use web search extensively. Label each finding:
- VERIFIED: Primary sources, official statements
- CREDIBLE: Multiple independent sources
- ALLEGED: Single source
- SPECULATIVE: Inference from patterns

Write findings to [output location].

No meta-context. No "previous iteration found X." Just the task.

Source Diversity in Worker Assignment

Critical: Assign workers sources from DIFFERENT categories to prevent perspective monoculture.

When spawning multiple blind workers on the same topic:

• Worker A gets Western mainstream + independent investigative sources
• Worker B gets non-Western/Global South + anti-interventionist sources
• Worker C gets financial/market + think tank sources (with funding noted)

This is the research equivalent of jury selection — different vantage points produce different signal.

See SOURCE_DIVERSITY_FRAMEWORK.md (in the cui-bono skill) for the full source taxonomy with 6 categories, outlet-level strengths/weaknesses, and think tank funding maps.

Why Blindness Matters

1. Prevents confirmation of previous conclusions
2. Fresh perspective on each pass
3. Avoids "groove deepening" (reinforcing existing patterns)
4. Context rot - LLM performance degrades as context fills
5. Source diversity - Different workers using different source categories prevents echo chamber

Layer 3: Dialectic Synthesis

Method: Thesis → Antithesis → Synthesis

For each major claim from worker research:

Step 1: State Thesis

What does the evidence suggest?

Step 2: Find Antithesis

What would DISPROVE this? What counter-evidence exists?

Step 3: Synthesize

What nuanced conclusion accounts for both thesis and antithesis?

Source Category Check

During synthesis, verify: Do thesis and antithesis come from the same source category?

If both thesis and antithesis come from, say, Western mainstream sources — you haven't actually tested the claim against a different vantage point. You've found disagreement within one perspective.

True dialectic requires sources positioned differently relative to power, geography, and economic interest. A thesis from the Financial Times and an antithesis from Al Jazeera or Responsible Statecraft tests the claim more rigorously than FT vs Bloomberg (same category, same vantage).

Example

Thesis: European illustrators can't access US children's book market. Antithesis: Some European illustrators DO succeed in US market. Synthesis: Market access is theoretically possible but structurally constrained through specific barriers (tax, agents, style). Success requires specific pathways (prizes, diversity initiatives).

Nested Loop Structure

Combine with iterative-loop-engine for multi-level investigation:

OUTER LOOP: Investigation completion
├── INNER LOOP 1: Define terms/scope
├── INNER LOOP 2: Region/category A research
├── INNER LOOP 3: Region/category B research
├── ...
├── INNER LOOP N: Cross-category synthesis
└── EXIT when all criteria pass

Each inner loop follows iterative-loop-engine criteria:

• Define completion criteria
• Execute passes until criteria pass
• Never claim completion with failing criteria

Evidence Tier System

Completion Criteria

Investigation is complete when:

• All planned inner loops executed
• Evidence labeling ≥80% of claims
• Independent sources ≥2 per major claim
• Dialectic synthesis completed for major findings
• Counter-arguments explicitly addressed
• Nuanced conclusion (not binary unless evidence is binary)
• Limitations admitted

Promise: ALL FALSIFICATION CRITERIA PASS

What Worked (From Actual Use)

Effective Patterns

1. Blind worker agents - Each produced comprehensive, unbiased research
2. Nested loop structure - Clear progression from definition → research → synthesis
3. Evidence tier labeling - Made verification explicit, prevented overclaiming
4. Dialectic synthesis - Forced consideration of counter-arguments
5. NO TOKEN LIMIT directive - Allowed thorough investigation

Areas for Improvement

1. Adversarial verification - Did not spawn dedicated falsification agents
2. Cross-source triangulation - Could structure explicit cross-validation
3. Orchestrator intervention logging - Should formally track prompt mutations
4. Quantitative gaps - Some claims rely on qualitative reports

See layers/improvement-notes.md for detailed analysis.

Integration Points

• iterative-loop-engine - Provides loop structure and criteria files
• deep-investigation-protocol - Provides evidence flow tracing and bias detection
• cui-bono (if available) - Provides power structure analysis for complex topics
• cui-bono/SOURCE_DIVERSITY_FRAMEWORK.md - Source taxonomy for multi-perspective worker assignment and dialectic source category checking
• frame-rotation - Can help escape stuck thinking patterns

Quick Start

1. Define what "done" looks like - Explicit criteria
2. Design inner loops - What regions/categories need separate research?
3. Spawn blind workers - Task-only prompts, no meta-context
4. Collect findings - Let workers write to designated outputs
5. Synthesize with dialectic - Thesis → Antithesis → Synthesis
6. Check criteria - All pass? Done. Gaps? Iterate.

Logging (CRITICAL for Improvement)

Public Logs (Git-tracked)

• Methodology evolution notes
• Anonymized patterns observed
• Improvement suggestions

Location: This skill directory

Private Logs (Gitignored)

• Full research transcripts with personal data
• Specific user queries
• Detailed intermediate outputs

Location: .claude/local/research-logs/

Reference private logs from public methodology notes WITHOUT including personal content.

Generalization Notes

Applies To (Tested)

• Multi-market research (cultural differences)
• Product/brand investigation (from deep-investigation-protocol)
• Factual accuracy verification

May Apply To (Untested - Experiment Carefully)

• Technical architecture decisions
• Code review with bias concerns
• Creative research (style exploration)

Probably Doesn't Apply To

• Simple Q&A
• Implementation tasks with clear specs
• Tasks without verification criteria

When uncertain: Try it, document results, update this section.