It doesn’t predict the market. It prices it.

An options desk that measures what movement costs against what movement actually is — and refuses any trade whose stress grid breaches its budget. Paper only, by construction.

Enter the desk Read the architecture

The SKEW desk: universe rail, volatility instruments, risk authority and the audit stream

Built onAlpaca Trading APIAlpaca Options DataModel Context ProtocolClaudePythonNext.js

What every other agent does

Predict direction. Read the headlines, or a moving average, or a model’s intuition, and buy an option pointing the way it guesses. The option is incidental — a leveraged bet on a forecast that neither the model nor anyone else can reliably make.

What this desk does

Measure what movement costs against what movement actually is. Implied volatility runs persistently above the volatility that gets realized, because people pay for protection. That gap — the variance risk premium — is structural, documented, and requires no forecast at all. Direction is never an input.

how it works — three steps, no forecast

Measure

Implied volatility against realized, across eight names. When the market charges more for movement than movement delivers, that gap is the signal. Direction is never an input.

Construct

Defined-risk structures sized backwards from the risk budget. Maximum loss is computed before the position exists.

Prove

Every structure is repriced across 84 scenarios before execution. One breach and it doesn't trade.

the instruments — every cell is the real desk

the premium, right now

Loading the recorded state — one beat, never a blank.

decisions traced

0 refused · 0 executed — every one replayable to its inputs

paper only

No live code path exists.

base url pinned to the paper endpoint
startup refuses anything else
defined-risk structures only

dedicated competition account
starting balance — awaiting first armed boot

realized-vol cone

loading…

term structure

loading…

last decision

No decision recorded yet — the log fills with the first cycle.

risk authority

loading…

Budgets grow only with clean closed trades; a drawdown demotes the tier automatically. There is no setting to raise them.

mcp surface · the desk as tools

scan_volatilityvol state for the universe
propose_structuresdefined-risk candidates
stress_testthe 84-scenario grid
risk_statustier, budgets, headroom
positionsopen book, marked
audit_logevery decision, with reasons
desk_statusarmed, market, account
executegated, confirm-required
closegated, confirm-required

Every read is open; the two mutating tools run the same gate chain and require explicit confirmation. Claude connects to this desk the same way you do.

The stress engine refuses any structure whose grid breaches the earned budget. No breach has been recorded yet — this exhibit fills in with the first real one, never with a mock.

deterministic gates · bounded selector

liquidityearningstermstressbudgetselectorexecute

Most cycles end here.

The model can choose among approved structures. It cannot invent one.

by the numbers

scenarios per candidate

deterministic gates

live code paths

decisions traced

questions a judge should ask

Is this real money?

No, and it cannot become real money by accident. The system runs on Alpaca’s paper API only: the base URL is pinned to the paper endpoint, startup asserts it and refuses to boot against anything else, and there is no live-trading code path anywhere in the repository — not even behind a flag.

What is the edge?

The variance risk premium: implied volatility runs persistently above the volatility that gets realized, because people pay for protection. The gap is structural and documented in the academic literature, and harvesting it requires no forecast of direction — which is why this desk never makes one.

What if the model fails or is unreachable?

The desk abstains, deterministically, and logs the abstention with the reason. The model’s only power is choosing among pre-validated candidates or declining them; it cannot invent contracts, change strikes, or bypass a gate. A selector outage therefore fails to no-trade, never to wrong-trade — and the startup preflight refuses to report the desk as armed while the selector is unreachable.

Why is there no IV rank?

Because Alpaca serves no historical implied volatility, and a 52-week IV rank computed from data that does not exist would be fabricated. The desk builds its own IV history forward from first run and refuses to print a rank until it holds twenty distinct trading days — until then it says exactly how many days it has collected.

How is maximum loss guaranteed?

Every structure is defined-risk by construction: a long option caps each short one, so the worst case is a known number, not an estimate. That number is computed when the structure is assembled, asserted against the per-trade budget at construction, re-checked by the budget gate, and stress-tested across 84 scenarios before any order exists.

Can I inspect a decision?

Yes — every decision, including every refusal, records its full chain: what was scanned, measured, classified, built, which gate stopped it and why. Here is a real refused trade’s trace — nothing on that page is recomputed for display.

See what it decided today.

Enter the desk