How to Prevent AI Hallucinations in Code ($0, local, 3ms)

AI coding tools like Cursor, Claude Code, and Cline are incredibly productive — until they hallucinate. They invent APIs that don't exist, reference functions with wrong signatures, and fabricate file paths. A single hallucinated import can cost hours of debugging.

WITNESS, built into Entroly, checks every LLM response against the actual source code before it reaches your editor. It runs locally, costs nothing, and catches hallucinations in 3 milliseconds.

The Hallucination Problem in AI Coding

AI hallucinations in code are different from hallucinations in general text. In coding, hallucinations manifest as:

• Invented APIs: The model calls response.getBody() when the actual method is response.content
• Wrong function signatures: Passing 3 arguments to a function that takes 2
• Fabricated file paths: Referencing src/utils/auth.ts when the file is actually lib/authentication.ts
• Non-existent imports: from sklearn.llm import Compressor — looks plausible, doesn't exist
• Outdated patterns: Using deprecated APIs from training data instead of current versions

These are hard to catch because they look right. The syntax is valid. The naming conventions match. But the code doesn't work.

How WITNESS Works

WITNESS uses Natural Language Inference (NLI) — a well-established NLP technique — to check whether claims in the LLM's output are entailed by (supported by) the evidence provided in the context.

1. Extract claims: WITNESS decomposes the LLM's response into individual factual claims
2. Match evidence: Each claim is matched against the source code files, documentation, and context that was provided to the model
3. Score entailment: Using STAVE (Statistical Textual Assessment and Verification Engine), each claim gets a support score
4. Flag or suppress: Unsupported claims are flagged (warn mode) or suppressed (strict mode) before reaching the developer

entroly witness --context-file evidence.txt --output-file answer.txt --mode strict

Comparison: Hallucination Detection Approaches

WITNESS achieves statistically comparable accuracy to GPT-4o-mini as a grounded judge — but at $0 cost and 250× lower latency.

Profiles for Different Workloads

WITNESS ships with tuned profiles for different use cases:

• code: Strict — fails closed on unsupported code claims. Best for auto-merge pipelines
• rag: Strict — for retrieval-augmented generation, suppresses unsupported retrievals
• qa: Strict — for question-answering, flags fabricated answers
• chat: Warn — for conversational use, warns but doesn't block
• summary: Warn — for summarization, flags unsupported inferences

entroly proxy --witness strict --witness-profile code    # suppress hallucinations in coding
entroly proxy --witness strict --witness-profile rag     # suppress in RAG pipelines

Setup: 30 Seconds

pip install entroly
cd /your/repo
entroly go

That's it. WITNESS automatically runs on every LLM response when Entroly is active. You can also use it standalone:

entroly witness --context-file evidence.txt --output-file answer.txt --mode strict

FAQ: AI Hallucination Prevention

pip install entroly && cd /your/repo && entroly go