Quality Check

The real problem: disconnected clusters

Individual orphan nodes are prevented by graph_batch (every node must connect). But the graph can still fragment into islands — clusters of well-connected nodes that don't talk to each other. Three regions about different aspects of the same project with no bridges between them means the understanding is siloed.

Use graph_analyze({ include: ["gaps", "bridges"] }) to detect this. When you find disconnected clusters:

1. Identify what the clusters are about
2. Ask: what's the conceptual bridge? A shared principle, a tension, a consequence?
3. Create bridging concept nodes with edges into both clusters

Ask the user to stop and think

Periodically prompt the user to pause and maintain the graph. Don't just run quality checks silently — surface what you find and ask the user to participate:

• "The graph has 3 clusters that aren't connected — want to think about how they relate?"
• "There are 5 open questions from earlier. Any of these resolved by what we've done since?"
• "You haven't revised any beliefs in a while — has anything shifted?"
• "The graph is getting dense in one area but sparse in another — should we explore the gap?"

Graph maintenance is a collaborative act, not just a background process.

Quality signals

• graph_thermostat — DIVERGE (explore) or CONVERGE (synthesize)
• graph_score — structural quality, target 70+
• graph_analyze — gaps, bridges, open questions, conflicts

Cybernetic sense-making

graph_analyze({ include: ["gaps", "bridges", "questions", "conflicts"] })

Thermostat

graph_thermostat({ mode }):

• DIVERGE — explore, generate variety, tolerate contradiction
• CONVERGE — synthesize, resolve tensions, commit to positions

Trigger audit protocol

Before closing a turn where you did substantive graph work, check which triggers you did NOT use. Run:

graph_find_by_trigger({ trigger: "question" })
graph_find_by_trigger({ trigger: "tension" })
graph_find_by_trigger({ trigger: "prediction" })

If any of these return empty and you've been doing real thinking, something is missing. Real understanding involves questions you can't answer yet, tensions you haven't resolved, and predictions about what comes next.

Diverge/converge action routing

When graph_thermostat returns a recommendation:

DIVERGE (explore, generate variety):

• Use graph_chaos() to inject serendipity — random connections that break tunnel vision
• Use graph_discover() for grounded exploration — finds related concepts you haven't connected yet
• Create serendipity trigger nodes for unexpected connections
• Tolerate contradiction — don't resolve tensions prematurely

CONVERGE (synthesize, resolve):

• Use graph_analyze({ include: ["gaps", "bridges", "questions", "conflicts"] }) to find what needs connecting
• Bridge disconnected clusters with synthesis nodes
• Resolve or reaffirm open questions
• Supersede stale beliefs
• Commit to positions on tensions

Coherence workflow

Deploy when the graph needs consolidation:

• Bridge disconnected clusters
• Supersede stale nodes
• Resolve or reaffirm open questions
• Add tension edges where conflicts exist
• Don't add new ideas — only clarify existing structure