ScholarNet AI

We auto-generate YouTube Shorts for our 4 products. One ran this week for ScholarNet (a study/tutoring tool) and the narration said: "I graded 10 essays in 60 seconds and caught 20 grammar errors." Problem: ScholarNet doesn't grade essays. It's an AI tutor — explains concepts, builds study plans, runs Brain Battles (live quiz competitions between universities). Grammar checking is what QuillBot does. The Short had already published. Caught it by extracting frames from the live YouTube video and reviewing them on a phone-shaped viewport. Slide 7s showed comparison cards literally listing "QuillBot · Grammarly · ProWritingAid" — Gamma had extrapolated from the narration's writing-tool seed material. Fix shipped same day: - New defense layer: `PRODUCT_CAPABILITY_DENYLIST` — a regex map of features each product DOES NOT have. Catches "graded N essays", "caught N grammar errors", "paraphrasing", "QuillBot/Grammarly comparison" in narration before the LLM can ship them. Three retry attempts with explicit "use only real capabilities" guidance, then a sanitizer swap. - Pruned 4 writing-tool-adjacent SEO tags ("literature review", "citation generator", "thesis writing") that were biasing the topic generator toward writing tools. This is a different defense from our existing marketing-fab regex (which catches `$X saved` and `N% conversion rate`). Marketing-fab patterns look for QUANTIFIED claims. Capability hallucinations are narrative claims — different shape, different gate. Smoke test at `_test-capability-gate.js`: 14 positive/negative cases across 4 products, 16 capability labels asserted against the categorization regex. Passes deterministically. Builder Q: how do you defend against an LLM inventing features your product doesn't have? Regex denylist feels brittle, but I haven't found a better belt-and-suspenders pattern yet. 🔗 scholar.0xpi.com