A transit authority's RAG-based schedule assistant scores well on BLEU and ROUGE against reference answers, yet riders keep reporting confidently wrong schedule times in its replies. What should the team add to its evaluation process to catch this?

BLEU and ROUGE measure n-gram overlap between a generated answer and a reference text, so they reward phrasing similarity, not truth; a response can borrow the reference's wording almost exactly while substituting a wrong time or route and still score well, because neither metric checks the response against ground-truth schedule data. A human evaluation layer closes that gap by having a reviewer (or a structured rubric) confirm the factual content of a sample of answers against the authoritative schedule source, catching hallucinations that lexical metrics are structurally blind to. Raising the BLEU/ROUGE threshold fails by concept because it just demands more of the same kind of overlap the assistant is already achieving; a hallucinated answer that happens to reuse reference wording will still pass a higher bar. Growing the reference set improves how the metric is computed, not what it's capable of detecting, so the blind spot persists at any scale. Swapping in a semantic-similarity score still compares generated text to reference text rather than to the actual current schedule, so a fluently wrong answer can remain semantically close to a wrong reference. Scope note: human review doesn't scale to every response, so sample strategically around high-traffic routes. Operational check: track the rate of human-flagged factual errors over time as a leading indicator separate from BLEU/ROUGE trends.