Dimensions — Multi-Model Analysis

Dimensions are the heart of Knowledge Tree's approach to understanding. For every node in the graph, each configured AI model independently analyzes the same fact base, producing its own perspective. By comparing these perspectives, the system reveals where genuine consensus exists and where model biases determine conclusions.

The multimodal principle

Knowledge Tree uses multiple AI models — Claude, Gemini, GPT, Grok, Llama, GLM, and others — not as knowledge sources, but as diverse reasoning engines. Each model:

1. Receives the same set of facts for a given node
2. Analyzes them independently (no access to other models' outputs)
3. Produces a dimension — a structured analysis with content, confidence score, and suggested concepts

The key insight: when models trained on different data, by different companies, with different architectures all reach the same conclusion from the same evidence — that conclusion is likely robust. When they diverge, the divergence itself is informative.

What a dimension contains

Field	Description
content	The model's analysis text, grounded in facts
confidence	A 0-1 score reflecting the model's certainty
suggested_concepts	Related topics the model recommends exploring
fact_count	Number of facts provided to the model
model_id	Which model generated this dimension
model_metadata	Token usage, generation parameters

Fact citations

Dimensions include inline citations to specific facts using markdown links: [brief description](/facts/<uuid>). This means every claim in a dimension can be traced to the specific fact that supports it, and from there to the original source.

Dimension generation process

1. Load all facts linked to the node
2. Load neighbor context — dimensions from connected nodes (via edges) provide additional context
3. For each configured model:
   • Send the same fact base + neighbor context
   • The model generates its analysis independently
   • Parse the response into a Dimension record
4. Compute convergence across all dimensions
5. Store all dimensions and the convergence report

Node-type-specific prompts

Different node types get different analysis prompts:

Node Type	Focus
Concept	What the evidence reveals, patterns, connections
Entity	Role, factual details, relationships, involvement
Event	Timeline, causes, effects, participants, context
Perspective	Build the strongest case, note challenges as obstacles

For perspective nodes specifically, facts are first stance-classified as supporting, challenging, or neutral. The dimension then presents the perspective's case using its supporting facts while acknowledging challenges.

Convergence

After all dimensions are generated, a convergence report is automatically computed:

Convergence score (0-1)

• > 0.7 — Strong consensus across models. The conclusion is well-supported.
• 0.4 - 0.7 — Moderate agreement. Some uncertainty remains.
• < 0.4 — Significant disagreement. Multiple competing interpretations exist.

What the report contains

Field	Description
convergence_score	Overall agreement level
converged_claims	Claims all models agree on
recommended_content	Synthesized view of the consensus
divergent_claims	Where models disagree, with each model's position and analysis

Divergences

When models disagree, each divergent claim records:

• The specific claim in question
• Each model's position on it
• The divergence type — whether models interpret the same facts differently, emphasize different facts, or reach different conclusions from the same reasoning
• An analysis of what might explain the divergence

Divergences are surfaced transparently to users. They represent genuine areas of uncertainty or bias — exactly the kind of information that a single-model system would hide.

Why multiple models matter

A single AI model can:

• Overweight certain types of evidence based on its training data
• Reflect the biases of its creators or training process
• Miss patterns that other architectures would catch
• Present confident-sounding conclusions that lack genuine support

By requiring convergence across diverse models analyzing the same evidence, Knowledge Tree provides a level of epistemic robustness that no single model can achieve. The system doesn't hide disagreement — it highlights it as valuable information.