Can three models argue their way to a better investment decision?

Retail investors face information overload, fragmented workflows, and too many opaque tools. ZenInvest is my open experiment in reducing that burden without pretending judgment can be outsourced: an investing system that brings screening, research, debate, and execution into one governed workflow.

The design bet is simple to state and hard to build: a committee of models that genuinely argues, wrapped in deterministic rules none of them can override.

A committee that actually argues

Each cycle starts by screening a tradable universe of 6,900+ US equities down to a focused opportunity set. Then the committee takes over: Claude argues strategy, GPT-4o plays the skeptic, and Gemini leads on risk. This is not ensemble averaging — the roles read and rebut each other, pulling live evidence through dynamic tool use across market data, news, filings, and macro research while they reason.

Above the debate sits a proactive macro layer that sets the market regime — RISK_ON, RISK_OFF, or NEUTRAL — for every cycle, so individual stock arguments always happen inside an explicit view of the wider market.

Rules no model can override

Whatever the committee concludes, deterministic Python has the last word. Concentration limits cap single names and sectors, a drawdown state machine steps through ACTIVE, CAUTIOUS, and HALTED, a policy-enforced cash floor holds, and the risk veto cannot be overridden by any LLM output. Under cost pressure the pipeline sheds models gracefully before it ever halts.

The system runs autonomously cycle to cycle, but every cycle is pausable and auditable, and chat-initiated trades require explicit human confirmation.

The models get a voice. They don’t get the keys.

Learning that stays behind glass

The most careful piece is the learning loop. Every decision and outcome becomes training data — but the loop is read-only. Nothing it learns can influence live trades until it clears alpha-adjusted, regime-stratified promotion gates. Improvement is earned against evidence, never assumed.

Why build it in public

ZenInvest works as three things at once: structured research support for people who want help thinking without surrendering judgment to a black box, an AI-native laboratory for adversarial agent design and governed execution, and an operator command center that unifies research, risk, decisions, and outcomes in one dashboard. Operationally it behaves like a product, not a demo — 1,341 automated tests with fail-closed evaluation gates, Docker Compose deployment, full audit trails, and daily and weekly reporting with cost visibility.

The deeper reason to build it openly is the pattern. Adversarial agent roles, a deterministic veto, and learning gated behind evidence is a shape that transfers well beyond investing — it recurs across the other projects on this site.

ZenInvest is not financial advice. It is an educational, research, and product experimentation project currently running on a paper-trading account. Human oversight remains essential, and any live use should be approached cautiously.