What it does
VERDICT splits legal analysis across four agents with separate responsibilities and separate views of the record. CLERK handles intake and citation. WITNESS holds the factual record. COUNSEL argues a position. JUDGE weighs the arguments against the cited record.
The separation is the mechanism, not an aesthetic. An agent that both argues a position and evaluates it will find its own argument persuasive.
In the code
# Each role receives a filtered view. COUNSEL never sees JUDGE's
# deliberation; JUDGE never sees COUNSEL's private reasoning, only the
# filed argument and the cited record.
ROLE_VIEWS = {
"CLERK": {"intake", "citations"},
"WITNESS": {"record"},
"COUNSEL": {"record", "citations", "brief"},
"JUDGE": {"record", "citations", "filed_arguments"},
}Illustrative shape. Enforcing this in the context assembly rather than the prompt is what makes it a property of the system instead of a request.
How this differs from the ordinary version
Separation of powers, implemented
Most "multi-agent" systems are one model prompted four ways, sharing one context. Here the contexts are genuinely disjoint, which is the only version of the idea that changes the output.
The record is anchored
Evidence is hash-anchored through CCP, so what WITNESS held at the time of the analysis can be verified afterwards.
In the field
Where it fits
Document-heavy, adversarial, citation-bound work: the domain where a fabricated citation is not an embarrassing bug but a sanctionable event. The architecture is built around making that failure structurally hard rather than statistically unlikely.
Questions
- Is this legal advice?
- No. It is a research instrument for analyzing documents and arguments, not a substitute for a licensed attorney.