PenCite: teaching an LLM to cite itself
The hardest part of an AI academic writing tool isn't the writing — it's making every claim traceable to a real source.
PenCite is my AI academic writing tool. The pitch is simple: it should never say something you can't verify.
The problem
An LLM is great at prose and terrible at provenance. Paste a draft and it will happily generate a bibliography that looks flawless and doesn't exist. For academic writing, that's disqualifying.
The approach
The core constraint: every generated claim must be anchored to a source you provide — a PDF, a paper, a note.
[user supplies sources]
↓
[PenCite writes draft]
↓
[every paragraph cites a highlighted span]
If a sentence can't be traced back to a highlighted passage, the tool flags it rather than faking a citation.
What's working
The provenance model is sound: authors bring their sources, PenCite writes within them. What's still being tuned is tone — academic register, hedging, and avoiding overclaiming.
It's a genuinely interesting constraint. I'll keep posting updates here.