hallucination-risk — Hallucination Risk Detection

Severity: WARN · Auto-fix: No · Category: 🏗️ Structure

What It Does

Detects prompts that ask for factual or current information without providing grounding context. Open-ended factual questions are the primary trigger for model hallucination — the model generates plausible-sounding but invented answers.

Trigger Conditions

The rule fires when:

1. The prompt contains a factual question pattern (any of the below), AND
2. The prompt does NOT contain a grounding indicator (any of the below)

Factual Question Patterns (trigger the check)

Pattern group	Examples
Factual wh-questions	what is, what are, who is, who are, when did, when was, where is, how many, how much
Currency/recency signals	currently, latest, recently, today, now, as of, up to date
Summary requests	tell me about, give me a list of, give me a summary of, give me an overview of

Grounding Indicators (suppress the check)

Indicator	Example
Template variable	{context},
<context> tag	<context>The following is the report...</context>
Context: label	Context: Here is the data...
Code fence	```json
"Given the following"	Given the following document...
"Based on the following/above/provided"	Based on the provided text...
"Using the data/information/context below/above"	Using the context below...

Examples

Triggers the rule

What are the latest developments in quantum computing?
Who is the current CEO of OpenAI?
How many parameters does GPT-4 have?

Finding:

[ WARN ] hallucination-risk (line -)
  Prompt requests factual/current information without grounding context.
  Consider adding a {context} variable or <context> section with source data.

Grounded — passes

Based on the provided research brief, what are the latest developments
in quantum computing mentioned?

<context>
{{RESEARCH_BRIEF}}
</context>

Has Based on the provided + <context> → no finding.

False Positives

Factual questions with known-static answers — "What is 2+2?" or "What is the capital of France?" will trigger the rule because they contain what is. These are legitimate zero-shot questions where the model's training data is reliable. The rule errs on the side of caution — you can disable it for prompts you're confident are low-hallucination-risk.

Prompts that ground via the system prompt — if your grounding context is in a separate system prompt and your user-turn prompt just asks "What are the latest developments?", the rule will fire on the user-turn text alone. You can disable the rule if you handle grounding at the orchestration layer.

"Currently" in instructions — "Currently, the user is on step 3" doesn't ask for factual recall, but currently will still trigger the check. Use At this point instead.

Configuration

rules:
  hallucination_risk: true   # or false to disable

Demote to INFO:

rules:
  hallucination_risk:
    enabled: true
    level: info

Mitigations

The rule flags — it doesn't fix. Standard mitigations:

1. RAG (Retrieval-Augmented Generation) — retrieve relevant chunks, inject as context
2. Explicit grounding — "Based only on the document below, answer..."
3. Uncertainty instruction — "If you don't know, say 'I don't know' rather than guessing"
4. Source citation requirement — "Cite the specific sentence from the context that supports your answer"

Example: grounding pattern

<role>You are a research analyst.</role>
<task>Answer the question based strictly on the provided research brief.</task>
<constraints>
- Do not use knowledge from outside the provided context
- If the answer is not in the context, say "Not mentioned in the provided brief"
- Quote the relevant sentence when possible
</constraints>
<context>
{{RESEARCH_BRIEF}}
</context>
<question>{{USER_QUESTION}}</question>