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

ZenInvest started with a design bet that sounds right: a committee of models that genuinely argues, wrapped in deterministic rules none of them can override. One model argued strategy, an independent model challenged assumptions, a third took a risk-first view, and the roles read and rebutted each other while pulling live evidence from market data, news, filings, and macro research. Deterministic Python kept the last word on every trade.

It ran live on a Trading 212 Practice (paper) account from February to September 2026. Then I did what most AI trading projects never do: I audited the outcome against a baseline and published it.

What six months showed

The committee made 278 realised trades. Its win rate was 37.8% against a break-even of about 57%, its 30-day expectancy was negative in 23 of 24 weekly checks, and the account lost money while the index rose. The committee approved 96% of what it was shown, so the debate added very little selection. The strongest signal in the data was not in any model’s argument at all: positions held longer than 12 days won 65% of the time, and those held under 12 days won 28%.

A pre-registered intervention in the summer fixed a real mechanism, but not the one that mattered. The full account, with figures rebuilt from committed fixtures, is in the working paper Where LLMs Actually Add Value.

Three models agreeing is not evidence. A baseline is.

What ZEN-2 changes

The redesign keeps what worked (deterministic risk rules, broker-truth reconciliation, audited research tools, and a complete decision trail) and replaces the decision core:

• The model never sees a random sample. A deterministic engine ranks the full universe first, and the LLM only reads the top of that ranking.
• Every position carries a falsifiable thesis. One analyst pass a day writes a memo with a catalyst, horizon, entry, size, invalidation level, and review date. Exits follow the thesis, not the clock.
• The agent must beat its own controls. An index arm and a rules-only arm run alongside it on the same universe and costs. A keep, scale, or stop rule is frozen before the first trade and applied six months after go-live.

Autonomous buying is paused while ZEN-2 is built; exits, stops, and reporting keep running.

Why publish a negative result

Because it is the useful part. Most claims about LLM value in decision-making are never tested against a simple baseline, and the ones that fail are rarely written up. The pattern that transfers beyond investing is not the committee; it is the discipline around it: deterministic controls at the boundary, a baseline for every claim, a decision rule fixed in advance, and the willingness to report what the evidence says.

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