Fictional Case Agent
Status: Failed.
Your muse
Build a legal case agent like meta-models/legal-agent-workshop. It reads a fictional dispute file, drafts a checklist, proves each item with exact page quotes, asks me for evidence when it is stuck, and files a memo I sign off. Show every turn live.
The plan: Fictional Case Agent
Prove every checklist item with page quotes before you sign the memo.
- Upload the fictional dispute file and start the run
- Watch every turn live as it drafts the checklist and matches exact page quotes
- Answer the evidence prompts when it gets stuck
- Review the memo draft and sign off to file it
See the full plan
Who it's for
Law students and workshop teams practicing on fictional disputes.
How it works
- Upload the fictional dispute file and start the run.
- Watch every turn live as it drafts the checklist and matches exact page quotes.
- Answer the evidence prompts when it gets stuck.
- Review the memo draft and sign off to file it.
Screens
- Intake. Upload the fictional file, set the dispute question, and start the live run.
- Live Run and Memo. Show every turn, checklist with quotes, evidence asks, and the memo signoff.
What goes in
- PDF: Fictional dispute file, up to about 30 pages and under 4.5 MB. Text pages work best.
- Text (optional): Short typed replies when the agent asks for missing evidence.
What comes out
- Text: Checklist where each item links to exact page quotes.
- Text: Live turn feed showing drafts, quote checks, and evidence asks.
- File: Final memo file after the user signs off.
Models
Model Used for Why Muse Wonder Read the fictional dispute file, draft the checklist, match exact page quotes, ask for missing evidence, and draft the signoff memo with a live turn log. Only model that reads long PDFs and text and returns JSON matching a schema for checklist and memo. SAM 3.1 Locate signatures, stamps and exhibit marks on scanned pages. Finds and outlines signature blocks and marks from short prompts so the agent can cite where each mark appears. Limits
- Works only on fictional practice disputes. It is not legal advice and needs human review and signoff.
- Built fresh from the workshop idea with new names and content. It cannot clone or open the original repo.
- Quotes need readable text pages. Blurry scans or heavy handwriting may fail and trigger evidence asks.
- Long live runs use a larger model budget than simple tools.
- Built with Muse Muse and is not an official Meta product.
From your inspo
The person gave no files, only a note. They want a legal case agent for fictional disputes that proves each checklist item with page quotes, asks for evidence, and files a memo for signoff with every turn shown live.
Takes
- The job of a case agent on a fictional dispute
- The flow of checklist then page quotes then evidence asks then signoff memo
- The idea of showing every turn live
Leaves out
- The original repo name, code, text, or branding
- Any real people, firms, cases, or seals
Cost per run: about $1.03 (est.). 150k input tokens at $4 per M is $0.60 plus 20k output tokens at $20 per M is $0.40 plus 10 SAM image checks at $0.0025 is $0.025 for a total of about $1.03.
This build didn't ship.
The sandbox setup step failed: This builder snapshot doesn't have the muse-app overlay. Make a new one with npm run snapshot before turning MUSE_APP_LOOK on.
A broken product is never shown as working. Muse Code's notes are in its steps.