ADDE: adversarial due-diligence engine

TL;DR

• The most literal version of “a council of models from different labs”: each provider runs its own agent swarm over a virtual data room, and a council with one representative per provider arbitrates.
• A risk claim that only one provider raised, without evidence, gets dropped. A red-team tier can reject the result and send the graph back around.
• I wrote the spec and directed OpenAI’s Codex agent through the build in one long session.

What I built

Each provider’s swarm works the data room department by department. A team captain distills the swarm’s majority-backed findings into a strict pydantic “platter” of risk claims. A Grand Council, with one representative per provider, compares the platters and drops single-provider claims that arrive without evidence. An external-evidence tier checks what’s left against SEC EDGAR filings, NIST NVD vulnerabilities and CourtListener dockets, and degrades gracefully when a source doesn’t answer.

Departments run in parallel through LiteLLM, with per-provider concurrency caps and exponential backoff. The container runs as non-root with the data room mounted read-only.

What the live run showed

The mock-mode test suite passed. The first live run across several providers exited cleanly, but it wasn’t a clean result. Every department logged provider errors (unsupported response formats, rate limits), not every department produced valuation figures, and the financial desk missed a receivable trap planted in the data room.

That’s the lesson I carried forward: voting between models only helps when the votes are real and the facts underneath are checked by code. Concord and AccountWard both put deterministic verification under the model layer.

Status

Explored. It passed in mock mode only, and the last change (per-provider semaphores) was never re-run live.