Modus Ops

Demo

About

Modus Ops is a production-grade, multi-agent AI system that investigates real-world crimes through an adversarial panel of four specialised agents. Each agent with a fixed, opposing role. All the agents argue, challenge, and adjudicate every case before committing a tamper proof verdict on-chain. The four agents are: Law Enforcement (investigator, evidence gatherer, escalation routing), Prosecutor (builds the strongest affirmative case, maps evidence to statute), Defense Counsel (red-teams the prosecution, surfaces reasonable doubt and procedural failures), and a presiding Judge (weighs the adversarial record, issues a structured verdict with a confidence score). Agents do not summarize, they argue. The Defense cannot concede without a motion. The Prosecutor cannot hedge. Every finding produced by Law Enforcement is mathematically verified using Benford's Law, velocity analysis, shell entity ratios before it reaches the adversarial layer. Financial crime is the primary launch vertical. The architecture is domain-agnostic: the same adversarial pipeline handles any crime where evidence, testimony, and legal statutes can be structured and argued. The verdict contains prosecution brief, defense memorandum, judicial ruling, evidence hashes and is committed onchain as an immutable chain-of-custody record.

How it's made

The agent pipeline is orchestrated in Python as a sequential adversarial loop: Law Enforcement runs first (evidence extraction + mathematical verification via a dedicated Python code interpreter running Benford's Law, velocity, and shell entity ratio checks), followed by Prosecutor and Defense Counsel running in parallel (adversarial brief generation), with the Judge consuming the full adversarial record to produce a structured verdict object. Each agent is powered by a custom LLM with a role-locked system prompt that enforces adversarial stance. Agent identity is registered onchain via ERC-8004 on Arbitrum One. four tokens encoding role, version, and jurisdictional scope. The verdict payload (confidence score, dissenting positions, evidence hash) is committed onchain as an immutable audit record. Persistent case memory runs through a family of MCPs Modgudr, Ilamcetcenni, these are open sourced with AGPLv3 license. However Ramanujan-MCP, a super specialised MCP server that handles complex math, including calculus, trigonometry, linear algebra and non-linear algebra (after the famed Indian mathematician, Ramanujan) that compresses extracted facts into token-efficient shorthand, grades them G1–G5 by evidential confidence, and enforces a Recall → Verify → Commit loop to prevent contradictions across agent turns. Frontend: Next.js 14 App Router, CSS Modules, static export via nginx behind Cloudflare. No UI framework. Licensed BUSL-1.1. However Ramanujan MCP will never be made public.