Build strategies in a conversation, then inspect every rule

What problem does this solve for a broker?

Most clients hold a workable trading idea they have never written down precisely enough to test. The gap between "buy when it breaks out with volume" and an executable specification is where the majority of attempts stop, and it is a gap no amount of charting UI closes.

For a broker, that gap shows up as engagement that never converts into activity. Clients open a strategy builder, meet a blank rule form, and leave. The builder was not the wrong feature; requiring a specification as the first input was the wrong starting point.

Closing it by translation rather than by prediction is the modest and defensible claim. Language models are genuinely good at turning an imprecise description into a precise structure. They are not good at forecasting prices, and nothing on this page suggests otherwise.

Why does the industry keep getting this wrong?

Because the tempting version is a system that answers "what should I trade". This is advice, is regulated, and puts the broker in a position no compliance function will approve. The useful version answers "what exactly did you mean", which is a language problem rather than a market prediction.

That distinction determines everything downstream. A system that outputs a recommendation has to be defended as a recommendation. A system that outputs a reviewable rule set the client authored, in structured form, is a specification tool, and the client remains the author of their own strategy.

How does the builder work?

A five-stage pipeline: natural-language request, guided resolution of the asset and contract, structured rules, validation against instrument and risk services, and a backtest-ready version that can be compared with others.

The builder does not accept an underspecified strategy and fill the gaps itself. It asks. When a user names an instrument, the interface confirms the resolved symbol; when they name a timeframe, it locks it and states what that implies for holding period. Confirmed inputs stay visible through later revisions, so a user editing a rule in week three can still see what they agreed in week one.

Six input groups, captured and validated through the workflow

A strategy that cannot satisfy the last group is not deployable, which is why eligibility is an input to creation rather than a surprise at activation.

Where does the AI stop and the deterministic system start?

AI-native interpretation structures the trading intent. Deterministic schemas, instrument services, validation rules and risk policies decide whether a strategy can proceed. The model proposes; the platform decides.

That boundary is the single most important design decision on this page, and it is what makes the feature approvable. A model cannot talk its way past a validation rule, cannot resolve an instrument that does not exist, and cannot widen a risk limit. If the schema refuses, the strategy does not advance regardless of how confidently it was described.

The rule that follows from it is worth stating plainly: the platform refuses unresolved derivatives rather than substituting the cash underlying. A user who asks for an option the catalogue cannot resolve gets a refusal, not a quietly different instrument.

How do users improve a strategy without rebuilding it?

Two paths. Change one rule manually and generate an isolated candidate, or ask Loop to compare a bounded set of candidates against one declared objective. Promotion of any candidate remains subject to review, deterministic validation and configured evidence requirements.

"Bounded" is carrying the weight in that sentence. An unbounded search over parameters is a machine for producing overfitted strategies, and it produces them faster than a human can evaluate them. Constraining the comparison to isolated candidates against a single declared objective keeps the exercise interpretable, and keeps a reviewer able to say why one version was promoted over another.