⚠️ LLM Confabulation Risk by Query Complexity Simulator

Simple Queries Answered from Dense Training Coverage

Placeholder: common facts are well represented in training data.

• 1 / 5: Query complexity (single-step lookup)
• ~96%: Training coverage (placeholder estimate)
• Low: Confabulation risk (placeholder estimate)
• Minimal: Verification needed (placeholder guidance)

Why simple queries are low risk

Placeholder: dense coverage means the model has seen this pattern many times.

What reliable grounding looks like

Placeholder: high-frequency facts sit well inside the model's reliable-knowledge region.

Moderate Complexity Introduces the First Knowledge Gaps

Placeholder: bridging between known facts starts to strain grounding.

• 2 / 5: Query complexity (placeholder)
• ~80%: Training coverage (placeholder estimate)
• Low–moderate: Confabulation risk (placeholder estimate)
• Light spot-check: Verification needed (placeholder guidance)

Bridging between known facts

Placeholder: the model interpolates between two well-known facts.

Early signs of strain

Placeholder: fluent phrasing can mask small factual gaps here.

Complex Multi-Hop Queries Chain Several Reasoning Steps

Placeholder: each extra hop adds a chance of an unsupported leap.

• 4 / 5: Query complexity (placeholder)
• ~65%: Training coverage (placeholder estimate)
• Moderate: Confabulation risk (placeholder estimate)
• Step-by-step review: Verification needed (placeholder guidance)

Chained reasoning steps

Placeholder: each hop compounds uncertainty from the previous one.

Where the chain breaks

Placeholder: the weakest link sets the reliability of the whole chain.

Rare and Obscure Topics Force Interpolation

Placeholder: sparse training data sharply raises confabulation risk.

• 4 / 5: Query complexity (placeholder)
• ~22%: Training coverage (placeholder estimate)
• High: Confabulation risk (placeholder estimate)
• Full source check: Verification needed (placeholder guidance)

Interpolation past the boundary

Placeholder: the model extrapolates beyond reliable coverage.

Fluency without grounding

Placeholder: confident tone does not imply factual grounding here.

Confabulation Rate Rises Non-Linearly with Complexity and Rarity

Placeholder: hardest, highest-stakes queries carry the most risk.

• 5 / 5: Query complexity (placeholder)
• ~8%: Training coverage (placeholder estimate)
• Critical: Confabulation risk (placeholder estimate)
• Mandatory expert review: Verification needed (placeholder guidance)

Why risk compounds non-linearly

Placeholder: complexity and rarity interact, not just add.

Placeholder: scrutiny should scale with exactly these queries.

Practical implication for clinical use

Placeholder: extra verification matters most where it is skipped most.