> caseagent
An agent that argues an arbitration, one filing at a time, and tests the result on the DecisionLayer simulator. Write your own side of a real dispute, or let it play both sides and see who wins.
caseagent simulate saas-outage --dry-run: four phases written by Claude Opus 5.5, with per-call token counts and cost.How it works
A case is a dispute (claim form, contract, exhibits) plus a thread: the filings made so far. The agent writes the next filing for whichever side is due, against the thread as it actually stands, trying to win for that side. Each filing goes through a private brief, a draft, and a cite-checking review.
| Phase | Filed by | What the agent writes |
|---|---|---|
| Claim | Claimant | Opening argument and remedies |
| Answer | Respondent | Responses, defenses, and a counterclaim if the record supports one |
| Reply | Claimant | Point-by-point reply, defenses to any counterclaim |
| Response | Respondent | Final response |
Try it
Requires uv. The fake model shows the mechanics with no keys and no cost.
# clone and set up git clone https://github.com/DecisionScienceResearch/caseagent.git cd caseagent && uv sync # fetch the sample disputes, then a free dry run uv run caseagent samples fetch --all uv run caseagent simulate saas-outage -m fake:demo --depth quick --dry-run # or a guided walkthrough uv run caseagent tour
Models come through OpenRouter, so one key covers Anthropic, OpenAI, Google and others. Compare models head to head on the same dispute, and swap sides to check the case isn't simply favoring one of them.