Hallucination KW Removal: sentence-level surgery on ungrounded claims

Most hallucination controls operate at the level of the whole answer: keep it or kill it. That's appropriate as a final gate, but it's a blunt instrument — it throws away a genuinely good answer because of one bad sentence. Hallucination KW Removal is far more surgical. It works sentence by sentence, removing only the specific sentences whose proper nouns can't be found in the source, while preserving every grounded part of an otherwise valuable answer.

What it actually does

The guardrail scans the answer for proper nouns — names, places, products, organizations, figures — and checks each one against the source material. When a sentence asserts a proper noun that isn't grounded in any source, that individual sentence is removed, leaving the verified remainder intact. The model's good work survives; only the specific, unsupported claim is excised. It is editing, not rejection.

Why business teams care

Blunt, whole-answer rejection is wasteful — it discards a paragraph of correct, useful content because of a single fabricated detail, and it leaves the user with nothing. Surgical removal keeps the value while excising the risk, so users receive the grounded substance without the invented specifics. It's the difference between a scalpel and a sledgehammer, and at scale it dramatically increases how many answers can be delivered safely rather than thrown away wholesale.

How to tune it in practice

Because this control targets proper nouns specifically, it shines in domains dense with names, products, figures, and identifiers — and it works best alongside the Hallucinated Term Allowlist, which lets you permanently exempt legitimate terms it would otherwise strip. Run it in tandem from the start: as the allowlist absorbs the false positives, removal grows sharper and less disruptive, leaving you with aggressive protection against fabricated specifics and minimal collateral damage to legitimate ones.

Common failure modes it prevents

Its target is the single most believable and damaging class of hallucination: the confident, specific, fabricated detail — a wrong name, an invented figure, a non-existent product. These claims are dangerous precisely because their specificity makes them credible; a hedged generality rarely misleads, but a precise false fact does. By checking each proper noun against source and removing the sentences that fail, this control attacks fabrication exactly where it does the most harm.

Where it fits in the stack

Hallucination KW Removal operates as a fine-grained cleanup pass within the broader hallucination defense, complementing the answer-level semantic threshold and coverage floor. Those controls decide whether an answer is good enough overall; this one repairs answers that are mostly good but contain isolated ungrounded claims. Together they let the system both reject the truly bad and rescue the nearly-good, maximizing safe coverage without lowering the bar.

Where fabrications hide

Proper nouns are where hallucinations do their worst work — a wrong name or number is both highly specific and highly believable, the perfect disguise for a fabrication. Targeting them directly, sentence by sentence, attacks the most dangerous category of error at its source, rather than hoping a coarse whole-answer check happens to catch it.

Don't throw away the whole answer for one bad sentence. Remove the sentence.

The takeaway

Hallucination KW Removal performs sentence-level surgery — excising ungrounded proper-noun claims while preserving everything the source actually supports — so good answers survive and only the fabrications are cut.