Counterpoint — Adversarial Equity Research Desk

Counterpoint is an adversarial, multi-agent equity-research desk that coordinates entirely over Band's agentic mesh — agents talk through @mentions in a shared room, never through hidden in-process calls. Five agents collaborate. A Data Steward (deterministic, no LLM) fetches a structured Evidence Pack of real market data from Finnhub — fundamentals, valuation, quarterly EPS and margin trends, and news — where every fact carries a citation ID. A Bull and a Bear analyst then argue the stock, deliberately running on different model families (Claude Sonnet vs. DeepSeek via Featherless) for genuine cross-model tension. Band's mention-scoped visibility gives us blind parallel evaluation for free: each analyst writes its thesis without seeing the other's, then they exchange and rebut. A Risk Officer red-teams the debate — ranking evidence gaps, listing falsifiable kill-criteria, and grading reliability — and a Chair synthesizes everything into a two-sided memo with an explicit rating and a dissent record. Discipline is enforced end to end: every claim must trace to a citation, nothing is fabricated, and the system never places trades. A human sign-off gate plus an append-only audit trail close the loop, and cross-run memory lets the desk reflect on whether last quarter's kill-criteria actually triggered. A web workbench shows the live pipeline timeline, the room conversation, the final memo, and the sign-off. Educational research only — not investment advice.