Zero wrong answers with the model unplugged

2026-07-20 · 1 min read

How much of your "AI product" is actually just a thin wrapper around an LLM API?

We wanted to prove that Epistom's semantic core is robust independently of the underlying language model. So, we ran a cross-model gauntlet. We ran our full Natural Language Query (NLQ) evaluation suite with the LLM completely disconnected, throttled, or quota-capped.

The results? The gate abstained rather than guessed.

That run is not a number we can re-prove from this tree. What the code does guarantee: missing concepts and below-bar confidence withhold rows, and hard abstention is the default. Deterministic tiers (verified queries; the Ontop SPARQL tier, which is off by default) keep answering without an LLM.

This proves our core architectural philosophy: the gate does the work, not the model. The LLM is just a highly capable interface layer. The actual understanding, validation, and security come from the semantic trust layer beneath it. When the model goes down, your enterprise shouldn't start generating hallucinations—it should gracefully degrade to deterministic safety.