# Key Term Penalty: don't drop the names that matter > Key Term Penalty docks answers that omit the key entities present in the source — catching subtle drift before it ships. **Category:** Hallucination Prevention **Author:** NeuralSeek Team · **Published:** June 9, 2026 **Canonical:** https://neuralseek.ai/ai-grounded/key-term-penalty **Section index:** https://neuralseek.ai/ai-grounded Hallucination isn't only about adding false information; it's also about silently dropping the right information. An answer that quietly omits a critical entity — a specific product name, a particular regulation, a named party to a contract — can be technically true and practically wrong, because the very detail the user needed never made it into the response. These omissions are especially dangerous precisely because they read as smooth and confident. Key Term Penalty is the control that catches them. ## What it actually does The guardrail identifies the key entities present in the source material and penalizes answers that fail to carry them through. If a response should mention a specific account type, a drug name, a regulation, or a legal clause that the source clearly establishes as central — but doesn't — it is scored lower and treated as weakly grounded. The penalty makes the absence of an important entity count against the answer, the same way a false addition would. ## Why business teams care Omissions are insidious because nothing about the surface of the answer reveals them. The response is fluent, well-structured, and confident; it simply summarized away the one detail that mattered. Key Term Penalty surfaces exactly these cases — the answers that are smooth but incomplete — protecting users from being misled by what was left out rather than what was put in. In regulated domains, a missing entity can change the meaning entirely. ## How to tune it in practice Calibrate the strictness to how much precision your domain demands. In settings where the exact entity is the whole point — which specific policy, which named drug, which regulation — apply a strong penalty so the assistant is pushed hard to carry those terms through. In more general settings, a lighter penalty nudges toward completeness without over-penalizing reasonable paraphrase. Watch for answers that feel correct but vague; that's the signature of omissions this control should be catching. ## Common failure modes it prevents The classic failure is the 'helpful summary that drops the point,' where the model compresses the source into something readable but loses the specific entity that made it actionable. Another is 'entity blur,' where the answer refers to something generically ('the relevant plan,' 'the applicable rule') instead of naming it — leaving the user to guess which one. Key Term Penalty pushes the assistant to be specific where specificity matters. ## Where it fits in the stack Key Term Penalty works alongside Term Penalty, Coverage Weight, and the semantic threshold as part of the grounding-fidelity layer. While the semantic threshold judges overall answer-to-source match, Key Term Penalty zooms in on the entities that carry disproportionate meaning, ensuring they survive even when the broader answer scores well. The two together catch both wholesale ungroundedness and surgical omission. ## Precision where it counts In regulated industries, the difference between two similarly named entities is often everything — two drugs, two account types, two regulations that differ by a single word. Penalizing missing key terms keeps the assistant precise about the nouns that carry the meaning, so it never abstracts away the exact detail a correct answer depends on. > The most dangerous omission is the one that reads perfectly fine. ## The takeaway Key Term Penalty ensures the entities that matter survive the trip from source to answer — catching the fluent, confident responses that quietly leave out the very point the user needed. --- From NeuralSeek's AI Grounded — practical, web-verified guidance on building governed, grounded enterprise AI. NeuralSeek is the model-agnostic, governed AI platform you own: any LLM (swap with no rebuild), your data in your own tenant (cloud or on-prem), 118 guardrails enforced before any action, one container that runs anywhere.