While working with founders on applied AI inside spectup, one practice consistently caught hallucinations early, forcing the uncomfortable but necessary stop decisions. We built a small golden question set made from real customer queries, then intentionally poisoned it with adversarial prompts that mixed correct context with tempting but wrong facts. The rule was simple, every answer had to cite retrieved sources line by line, or it failed. That discipline alone surfaced issues no demo ever showed. I remember one engagement where the model sounded confident, fluent, and completely wrong once every ten answers. The grounding score we tracked was answer tokens supported by retrieved text divided by total answer tokens, and it hovered around seventy percent. On paper that looked acceptable, but when we ran the adversarial set, failure clustered around edge cases investors would absolutely ask about. That was the first time I personally blocked a launch despite product pressure. The metric that changed the go no go decision was attribution coverage on the golden set, not latency or accuracy averages. We required ninety five percent of factual sentences to trace back to retrieved documents, otherwise it was a hard no. We used a lightweight evaluation setup inside our existing stack, nothing exotic, but the clarity it gave the team was refreshing. One of our team members even joked it saved us from becoming confidently wrong at scale. After remediation, we narrowed retrieval scope, cleaned document chunking, and forced abstention when confidence dropped. The grounding score moved above ninety percent, and more importantly, failures became obvious refusals instead of smooth hallucinations. From a capital advisory perspective at spectup, this matters because AI risk shows up in diligence fast. Catching it early protects trust, product credibility, and ultimately the funding story.