Domain-Adaptive Grounding

The monks need deep grounding before they can believe effectively. But what constitutes grounding depends on the domain type and how novel it is. The skill must adapt.

​

Research Depth: The Main Calibration Knob

Research depth is the main knob. It’s the only phase that meaningfully changes the time and cost profile — everything else (essays, analysis, synthesis, validation, auditor) is fast regardless.

​

Novel/Obscure Domain

Examples: Emerging technology, niche policy, unfamiliar institution Strategy: Full parallel research — 2-3 agents, 150-250K tokens Why: The orchestrator’s training data is thin or outdated. You need the research to:

• Write good framing corrections
• Identify degenerate framings (the obvious, shallow version of the dialectic that won’t produce insight)
• Ground the briefing in specifics

​

Well-Known Domain

Examples: React vs Vue, microservices vs monolith, common career decisions Strategy: Skip or minimize research Why: The orchestrator’s training data is rich. Write the briefing from your own knowledge, perhaps with 2-3 targeted searches to check for recent developments. Savings: 10-20 minutes and 150K+ tokens

​

Known Domain, Novel Angle

Examples: “React vs Vue but specifically: how does OSS funding structure causally shape innovation character?” Strategy: Light research — a few targeted searches on the specific angle, not broad domain surveys Why: The orchestrator knows the landscape but needs to check the specific thesis.

​

Domain Types and Grounding Strategies

​

External-Research Domains

Examples: Engineering, strategy, policy, technical architecture These domains have literature, case studies, data, and named thinkers. The grounding comes from outside the user.

​

When Full Research is Needed

Run 2-3 parallel research subagents on different aspects of the domain. A natural split that works well:

1. Side A’s strongest literature — the key thinkers, evidence, and arguments for one position
2. Side B’s strongest literature — same for the other side
3. Broader landscape/context — institutional structures, historical parallels, adjacent domains, empirical data

Cost: ~150-250K tokens across agents Value: This is expensive but is the single most valuable investment in the entire process — deep grounding is what makes everything downstream good.

​

Personal and Values Domains

Examples: Life decisions, career, relationships, commitments, priorities These domains have little useful external literature. The grounding comes from the user themselves — their history, values, constraints, relationships, and patterns. The elenctic probing (Phase 1c) must go deeper and wider for these domains. You need to map:

​

The Full Landscape of Commitments

Not just the two in tension — everything the user is carrying. Ask: “Walk me through what’s on your plate right now — all of it.” Undifferentiated care (the Empathic Integrator pattern) only becomes visible when you see the full load.

​

The History

“Have you faced a decision like this before? What happened? What did you choose? How did it feel afterward?” The Exploratory Debater’s commitment pattern only becomes visible across multiple instances. The Practical Executor’s optimization lock only shows when you see what they haven’t questioned.

​

The Stakeholders and Their Actual Capacities

“Who else is affected by this? What can they actually do — not ideally, but right now?” This separates the vision from the reality, which is the Empathic Integrator’s core split.

​

The Values Underneath the Positions

“You say you value X and also Y. If you could only have one — gun to your head — which?” This surfaces the Possibility Explorer’s values hierarchy that they resist articulating.

​

The Constraints They’re Treating as Fixed

“What would you do if [constraint] disappeared tomorrow?” This reveals which constraints are real and which are assumed.

​

Limited External Research May Still Help

Search for frameworks, not facts:

• “How do people navigate career transitions at [user’s life stage]”
• “Decision frameworks for competing values”
• “What does research say about [specific situation type]”

This gives the monks structural scaffolding, not positions to believe — the positions come from the user’s own material.

​

Mixed Domains

Examples: Normative/institutional, creative direction These need both. A dialectic about institutional identity, for example, requires:

• External research (organizational history, governance structures, comparable institutions)
• The user’s personal values and judgment about what the institution should become

The interview needs to surface the personal dimension while the research agents cover the external.

​

What You Need to Know

In all cases, you need to know the domain well enough to:

1. Identify and correct likely degenerate framings (the obvious/boring version of the dialectic that won’t produce insight)
2. Generate specific research directives or interview questions for each subagent
3. Write framing corrections that steer monks away from shallow takes
4. Identify the deepest available contradiction

​

The Context Briefing Document

Synthesize everything — external research AND user-sourced material — into a single neutral briefing document and save it to a file (e.g., context_briefing.md).

​

For External-Research Domains

Cover:

• Key evidence, sources, and arguments from all sides
• The landscape of the debate — who the key thinkers are, what positions exist
• Relevant empirical data, historical context, institutional structures
• The specific framing you’ve identified as the deepest contradiction

​

For Personal/Values Domains

Cover:

• The user’s full commitment landscape (all the things they’re carrying)
• Relevant history and patterns (past decisions, outcomes, recurring themes)
• Stakeholders and their actual capacities
• The values hierarchy as best you can reconstruct it
• Constraints (which are real, which are assumed)
• The specific tension you’ve identified as the deepest contradiction

​

For Mixed Domains

Include both sections, clearly marked. Both monks will read this file before writing.