← Live dashboard
ARN Options · How it decides

The Cortex

Nine layers of judgment stand between market noise and a single options trade. This is the exact path a decision travels — the same logic running live, every fifteen seconds, on a real paper-trading account.

Scroll
01 The Filter

Most cycles, it does nothing — on purpose.

All 64 watched symbols are re-evaluated every 15 seconds. But before a single model runs, hard rules can veto a symbol instantly. Restraint is the first feature.

Scan · 64 symbols every 15s
0 symbols cleared the gate this cycle
TREND_UP CHOPPY TREND_DOWN
watches SPY watches VIX
02 Reading the Weather

First, it reads the whole room.

Before looking at any single stock, the algorithm classifies the entire market — recomputed every 30 minutes from SPY and VIX — into one of three regimes.

TREND_UP SPY grinding higher with consistent momentum, VIX calm.
TREND_DOWN SPY grinding lower with consistent momentum, VIX calm.
CHOPPY No clear direction, or VIX above 35. Options trading is paused entirely, system-wide — no matter how good a single stock looks.
03 Gathering Evidence

Three independent witnesses.

For every symbol that survives the filter, three separate signals report in — each blind to the others.

TO
MODEL
04 The Brain

Thirteen signals, one conviction.

Two gradient-boosted models score every symbol's odds of a meaningful move — and as of July 2026, they hold no veto over live trades. Forensics showed the trade type they gated couldn't clear its own costs, so the models were benched along with it. They now live in the Lab (Stage 09), fighting nightly duels on data they've never seen — a comeback requires a sustained out-of-sample record, not a good week.

Up model
0.00
conviction · logged, not obeyed
Down model
— off
benched · duels nightly
05 The Gate

Every layer has to agree at once.

The flagship trade — a sold put spread on SPY or QQQ — only fires when all six conditions click open simultaneously. Most days, at least one stays shut. (A parallel equity leg ran alongside this through mid-2026; retired 2026-07-28 after it never cleared breakeven across hundreds of trades. One strategy, proven, beats two, one of which quietly loses.)

Why sell premium instead of buying it?

A bought option needs the market to surge before time decay and spread costs eat it — 1,495 trades proved that math (Stage 09). A sold put spread wins in three of four endings: market up, flat, or even slightly down. The seller doesn't need to predict the rally — only that the crash stays home — and is paid up front for a risk that's capped to the dollar.

FIRE · SELL PUT SPREAD
Strike ladder · SPY puts, 10 DTE
SPY 748/743 bull put · credit $1.01
max loss $399 · defined at entry · one MLEG limit at mid
06 Pulling the Trigger

Two strikes, sold as one order.

Only SPY and QQQ qualify — the two chains liquid enough that the exit door is always penny-wide. The system sells the put sitting about one percent below the market and buys another five dollars lower as insurance on the insurance. Worst case is fixed to the dollar before the order leaves the building.

07 Letting it Breathe

Held with patience, closed with rules.

No panic exits — and no greed either. A sold spread's final dollars are its most dangerous: the position is bought back once 60% of the credit is captured, never ridden into expiration. Equities keep their intraday rules and sleep in cash; only the spreads hold overnight.

TAKE PROFIT · 60% OF CREDIT CAPTURED
STOP · DEBIT 2.5× CREDIT
holds overnight — by design
OUT BY 2 DTE · ALWAYS
60%
Credit captured → buy back
2.5×
Credit — hard stop
≤2 DTE
Never into expiry week
1 strategy
Spreads-only since 2026-07-28
08 Showing its Work

Nothing is a black box.

Every decision is logged with the exact 13 input values at that instant — plus SHAP attribution, showing precisely how much each input pushed the model's confidence up or down.

Every scan of all 64 symbols and every trade's full lifecycle land in a public database; receipts like this one accompany any model-gated entry — the standard every future strategy must meet before it goes live.

NVDA · CALL entry
decision receipt · 11:42:15 ET
0.68
confidence
SHAP attribution base 0.50 → 0.68
09 The Lab

New brains must beat the old one — on data neither has seen.

The live model is a standing champion, not a finished product. Challenger models are trained on richer pattern features — the shape of the whole day, not just the latest bar: streaks, range compression, position against the morning's range, how long since the day's high, how many times price has chopped across VWAP.

Then comes the duel: champion and challenger score weeks of market data neither has ever seen — re-run automatically every night. No clear win, no deployment; a tie goes to the incumbent. The same discipline killed an entire strategy candidate in one evening: swing-traded calls backtested at +3.3% per trade… until random entries under the same market filter scored +4.0%. All beta, no alpha, zero dollars spent learning it.

Next: learning from its own trades

Every decision receipt (Stage 08) is also training data. Once enough accumulate, a second model learns which of the bot's own signals actually made money — net of spreads — and vetoes the ones that historically haven't. The strategy starts studying itself.

Challenger trial · July 2026
holdout: 2 unseen weeks
Champion — live model
13 snapshot signals
31.0%
Challenger — pattern model
25 shape features · calibrated confidence
40.5%
hit rate on each model's 200 most-confident never-seen bars — the challenger led only at the extreme, and by less than the noise floor
SPLIT DECISION · NOT DEPLOYED

The trial also caught the champion cheating: its first "win" came from having trained on the test window. Retrained clean, its score fell from 49% to 31%. Every future duel is leak-checked by construction.

And then it does it all again — in fifteen seconds.

Nine layers, 64 symbols, all session long — and overnight, now that every position's worst case is fixed to the dollar before entry. Bullish only, defined risk only. No real money — every trade runs on a paper account, fully logged and replayable.

Open the live dashboard PAPER TRADING Powered by Alpaca
ARN Options