Quill

AI collaborator — legal drafting with citation verification. Per-span signatures; catches fabricated citations under ABA Model Rule 3.3.

What this proves

LLM-fabricated citations are the lawyer-disbarment failure mode (Mata v. Avianca, 2023). The fix isn't a more careful prompt — it's a per-span verification step that hits an external authority and either confirms or marks the citation as distinguishable or fabricated. The contract stores the verdict per citation, not per document.

What the agent reads

• The user's draft text + cited cases
• CourtListener REST v4 (/search/?type=o&citation=<reporter-triple>) — open endpoint, no auth
• Bluebook citation strings parsed client-side to extract the reporter triple

Decision

Per span: VERIFIED (CourtListener has the case + correct citation), DISTINGUISHABLE (case exists but the cited proposition isn't supported), or FABRICATED (no matching case in CourtListener). Each span's verdict gets its own reasonHash write.

Code map

• Contract: contracts/src/QuillCoAuthor.sol
• CourtListener client: ui/src/lib/courtlistener.ts
• Verify route: ui/src/app/api/legal/verify/route.ts
• LLM call: ui/src/app/api/demo/[slug]/route.ts (quillHandler)
• UI: ui/src/app/quill/page.tsx, ui/src/components/poa/QuillDemo.tsx

Try it

demo-agents.theseus.network/quill. Paste a brief or pick a preset with deliberately mixed real + fabricated citations. The agent inline-marks each span with its verdict and the CourtListener URL.

On-chain

Per-span verdicts go to QuillCoAuthor. touchedIdCount() increments per verification.

---

Last updated: June 17, 2026.