In options, a bad estimate becomes a funded loss. So every number we publish is cross-validated, calibrated against outcomes, and gated by guardrails that refuse suspect output. Here's how.
Constraints, not features. Every recommendation passes all six.
Every American-style price is computed by two independent models — a closed-form analytical engine and a binomial-tree validator. Within 1%, we publish. Up to 5%, the conservative model wins. Beyond 5%, the candidate is dropped.
Every chain is checked against four classical invariants before any strategy logic runs. Violations become a data-quality warning on the trade card — never silently used.
Held positions: live chain delta — the market's own estimate. Scored candidates: a four-factor composite (path sim, early-exercise, dividend, event), surfaced with a transparent model spread. Per-leg delta, gamma, vega, theta on every card — in calendar-day convention, so theta matches your broker's Friday-to-Monday tape.
Five live regimes, classified from spot vol, term structure, skew, and realized vol. Strategies that don't belong in the current regime are blocked — selling premium in a crisis takes an explicit override.
Every published probability is scored against the realized outcome. Rolling Brier and Expected Calibration Error, broken down by strategy and DTE. Drift is visible; persistent drift triggers a re-tune.
Day-over-day P&L is decomposed into delta, gamma, vega, theta, and a residual. A growing residual means the model has stopped explaining reality — and we treat it as a flag, not a footnote.
Nothing is ranked before it has passed every stage.
Quotes and chains normalized, time-stamped, structurally validated.
Monotonicity, butterfly, calendar, put-call parity. Failures tagged, not used.
Two independent models per leg. Disagreement measured; outliers dropped.
Liquidity, regime, and expected-value tests gate what's even rankable.
Multi-factor score with risk-style overlay. Rationale and warnings attached.
Fail any gate, you don't appear — not even with caveats. A quiet screen beats a confident recommendation built on a stale or mis-priced quote.
Concentration, tax efficiency, and data freshness — built into the engine, not optional settings.
Per-trade caps, a 30% sector ceiling, and per-strategy limits are enforced on every candidate. In stressed markets, a volatility-regime multiplier scales position size down automatically.
Closing a position triggers lot ranking by short- vs long-term gain impact. FIFO, HIFO, and specific-ID methods are supported; the wash-sale window is respected.
Live feed during market hours, explicit TTLs per data type, automatic fallback when an upstream fails. Every number on screen has a known maximum age.
Pull the AI layer out tomorrow and every number on the platform is unchanged.
Isolation, encryption, observability — the boring parts where most platforms quietly fail.
Not whether the code runs. Whether the financial math holds.
Probabilities outside [0, 1]. Annualized returns exploding at 1 DTE. Spread payoffs going imaginary at the extremes. Each has a test — usually written the day we caught it.
Every supported strategy has dedicated tests asserting its financial properties hold under adversarial inputs.
Each statement format has its own regression suite. A change to one parser cannot silently break another.
Tests run against a real database before any release. Math-touching releases require a clean run on production-shaped data.
Enforced on every change — not aspirational.
If a model can't price, fit, or fetch — the trade card says so. Silent fallbacks erode trust.
When two models disagree, we show the spread. We don't pick a winner and pretend the other never ran.
Biased toward rejecting candidates. A blank screen is an acceptable outcome; a confident wrong recommendation is not.
Decision support, not auto-execution. Every order is yours to review and submit. We never move money on your behalf.
Robustness is the result of disciplined choices, repeated on every release.
Request Beta Access → Back to home