There’s a particular way language models fail that doesn’t look like failure at all.

The response is fluent. Grammatically perfect. Confident. It sounds like it knows what it’s talking about. It uses the right vocabulary, the appropriate tone, the expected structure.

But somewhere between the question and the answer, something didn’t happen. The model didn’t do the work.

Here’s the setup: you have a RAG system. User asks a question. Your system retrieves relevant documents. The model gets the question and the evidence. Everything it needs to answer correctly is right there in the context window.

And yet the answer is wrong. Not because the retrieval failed. Not because the documents were bad. The evidence was there. The model just didn’t use it.

We call this semantic laziness.

Geometry as measurement

A grounded response has a specific relationship to its source material. A lazy response has a different relationship — it stays closer to the question, orbiting the query without traveling to the evidence.

This relationship is measurable. Not by asking another model. By geometry.

Grounded responses move toward their sources. Lazy responses don’t. That difference leaves a signature you can detect.

Triage, not truth

Flag the lazy responses. Let the grounded ones through. Review the uncertain cases.

Not perfect. But better than hoping.

---

This concept is formalized in the Semantic Grounding Index (arXiv:2512.13771)

---

← Back to blog