Most trading products hide how they work and show you a curve that goes up. We do the opposite: here is the full reasoning pipeline every trade passes through — and the honest reason we publish the architecture but not the calibration.
When you describe a strategy in plain English, STRIX compiles it into a precise configuration: what signal type it trades (momentum, mean reversion, breakout), what it's allowed to buy, how much it risks per position, how it takes profits and cuts losses, and what timeframe it lives on. Every decision downstream is made inside those boundaries. Two users' strategies never get silently merged into one house algorithm — your strategy keeps its identity.
Each scan cycle, the engine sweeps the market for candidates that fit your strategy's signal — intraday strategies are evaluated on intraday behavior, swing strategies on multi-day behavior. A signal firing is the beginning of scrutiny, not the end. Candidates are ranked, and only the strongest survive to the expensive stages.
Every candidate then has to clear a series of independent quality gates. Each one exists because a real pattern of losing trades in our ledgers demanded it:
Options with poor two-sided quotes are rejected outright — a wide bid-ask spread means paying a built-in loss on entry and exit, before the thesis even gets a vote. Cheap contracts face the strictest version of this gate, because that's where spread costs hide best.
The engine estimates what an option should cost from the stock's actual recent volatility. When premium is trading far above that — the classic post-spike inflation — the trade is skipped. Paying double fair value means the move has to continue just to break even.
A move that has already run a long way is statistically more likely to snap back than to keep going. The engine refuses entries that chase over-extended moves — a filter written directly from our own losing trades, which clustered exactly there.
Cross-strategy checks stop your account from quietly stacking into one big directional bet, one sector, or one symbol held by another of your strategies. Diversification is enforced at the book level, not assumed.
A strategy that demonstrates it's losing over a meaningful sample is automatically blocked from opening new positions until things change — with a notification telling you why. The system would rather trade less than keep funding a configuration that doesn't work.
The last gate before execution is an AI review of the specific trade: does this contract, at this price, in this context, contradict the strategy's own stated rules? The AI cannot originate trades, cannot override your risk settings, and cannot say anything but ENTER or SKIP with a reason. It's a pre-flight inspector, not a pilot — and its reasoning is logged on every trade so you can audit it.
Position size comes from your risk budget, then shrinks for hostile conditions: more volatile underlyings get smaller positions so that one gap doesn't define your month, and platform-wide learning can scale risk down when recent conditions have been unkind. Size is never increased beyond your configured budget.
Every cycle, the engine manages existing positions before hunting for new ones — protecting money you have always outranks deploying money you haven't. The exit ladder includes trailing stops that let winners keep running instead of amputating them at a fixed target, stop logic watching the underlying stock (which moves more honestly than option premium), time-based exits that refuse to sit through the worst of option decay, and a hard rule that intraday strategies go flat before the close — gap risk isn't a scalper's job. Every exit is labeled with its reason in your ledger.
Every closed trade across all strategies feeds a collective learning layer. It learns transferable lessons — what conditions favor which behavior — never "copy this user's picks." Its influence is bounded by confidence math (small samples are treated as noise until they aren't), decays with time so dead patterns retire, is continuously A/B-tested against a control group of trades it doesn't touch, and is wrapped in a circuit breaker that de-risks everything when recent results degrade. When the learning layer adjusts your strategy's behavior, it nudges within your configuration — it never replaces it.
Everything above describes the architecture, and we publish it because you should never trust a black box with trading decisions. What we don't publish: the thresholds, formulas, weights, and prompt engineering behind each gate — values calibrated continuously against thousands of live paper trades and re-learned daily. An architecture diagram doesn't trade. Calibration does, and that's the part years of live ledgers bought us.