LLM Hallucination Debugger

Illustrative Simulation — not real model probabilities

Next-token prediction samples from a probability distribution, so some uncertainty is inherent. But many hallucinations are reducible system failures: missing retrieval, no verification, no abstention policy. This simulator shows how mitigations lower risk without ever reaching zero. All numbers are deterministic illustrations, not measurements.

Task Type
Signals
noneweakmoderatestrong
Interventions

Token-Branch Confidence

Generated Answer simulated

Evidence coverage
Mean claim confidence

Ungrounded vs Grounded Workflow

The left panel ignores retrieval and verification; the right applies your interventions. Risk updates instantly as controls change.

Ungrounded Model

Avg claim confidence
Evidence coverage

Grounded Workflow

Avg claim confidence
Evidence coverage

Claim–Evidence Matrix

Select a claim row to inspect simulated supporting snippets, contradictions, and the recommended action.

Knowledge Check

Why this matters

A language model does not look facts up by default; it continues text with plausible tokens. When context is thin and temperature is high, fluent-but-unsupported claims become likely. Retrieval grounds claims in documents, tools handle arithmetic and lookups, cross-checking catches contradictions, and abstention policies convert low confidence into honest uncertainty. These mitigations reduce risk substantially — they do not eliminate it, and calibrated humility remains part of good system design. High confidence is not proof: calibration means confidence should track accuracy, and it often does not without deliberate engineering.