A row of canvas tents along a high plain at golden hour, a bell tower on the ridge

mostly harmless markets presents

GRIDFLY

Don't predict — we sell the day. Five at the table, one blade in the way. Hold to the bell, let the premium decay.

GridFly is a structured options-income strategy wrapped in AI agent risk control. Built for the Alpaca AI Trading Agents Hackathon on the Alpaca MCP server, with Claude in the decision loop: it proposes and journals, and it never touches the risk path. hypothesis → controls → council → oracle → measurement → live. Six chapters, one week, in public.

in association with Alpaca

Kalshi

chapter one · the base hypothesis

THE GRAVEYARD

Two years, fifty strategies, one machine that kept its word.

rows of stone markers each draped with a collapsed tent

Two years in the graveyard. Fifty strategies buried, every one with a beautiful backtest. Forty-four grand in the river, and the river made no sound.

a workshop table with glowing slabs, a crystal ball, and one plain wooden box

Only one machine kept its word: the box that wrote everything down. Nearly two years of 0DTE option chains, recorded live through Alpaca's OPRA feed. The backtest corpus is our own archive, not vendor history.

a lone figure turning from the graveyard toward a bell tower, carrying a tent pole

Then the lesson: change the thing you trade, not the way you predict. Sell defined-risk SPX structures on a fixed intraday schedule and hold them to cash settlement. No direction, no forecast.

paper lanterns carried down a dark river

The hypothesis: option buyers systematically overpay for protection, and a disciplined seller collects that premium. The strategy is at its best on boring days. So the agent's real job is deciding, every half hour, whether the day has stopped being boring.

The naked hypothesis, backtested — no council, no gates, no cuts. These four stones exist to show why everything below exists. Backtest figures assume a fixed $100k redeployed daily so days stay comparable. The competition account deploys full equity every day, so gains compound into the next day's size: a competition posture, not a recommendation.

Win rate · raw

–

share of green sessions, before any controls

Mean day · raw

–

on $100k fully deployed

Worst day · raw

–

this is why the controls exist

TAIL

Sessions

–

0DTE chain history, entries at real quotes

The naked hypothesis makes money on average and occasionally loses a catastrophic amount. Everything below is about deleting the catastrophe without deleting the average.

chapter two · how it works, and the controls

THE TENT

A pole where the price is, two stakes to hold the frame, and the bell at four.

a standing stone at the camp entrance with a carved line, a brass gauge, and a wind ribbon

Gates are code with numbers frozen before the session, carved in stone at the camp entrance. The model gets a say. The risk path can't be crossed. Hope was never in the play.

THE POLE = where the price is.

TWO STAKES = the wings. Max loss is known before you enter — that's the only place to hide.

THE CANVAS = the credit. Cash-settled by the tower at 4:00. No delivery, no swerve.

The full configuration, backtested — the council plus every control, same sessions, same sizing.

Win rate · days traded

–

of the 134 days it chose to trade; it sat out the other 25 entirely

Mean day

–

on $100k fully deployed

Worst day

–

vs the naked hypothesis

Worst 10% of days

–

CVaR-10: the average across the worst tenth of all days. The tail number risk desks actually watch.

BUST

Compounded · raw

redeploying full equity daily, the naked hypothesis goes bust: one tail day exceeds the account

STANDS

Compounded · full config

survives

every one of 159 sessions, worst stretch included. Survival is the product; everything else is scale

Count every day, including the 25 it refused to trade, and the win rate reads like a coin flip. That is by design: the council trades hit rate for tail safety on purpose, and on the days it does deploy, its wins run far larger than its losses. Backtests use real historical option quotes with conservative exit pricing and modeled fees, on a proxy corpus of the same structure geometry. Backtests are a lab, not a promise. The measurement in chapter five is the experiment.

chapter three · the agent in the room

THE COUNCIL

Every half hour, one question: is today still boring enough to sell?

five silhouettes at a long curved table before a glowing doorway

Five seats, but they ain't equal. Weight is something that you earn, and the only way to earn it is to survive a pre-registered study. Two can vote and three just talk — and the talkers keep a score.

DECIDES

THE SCHOLARrealized volatility · the decisive seat

Calm, I let 'em sell it. Fast, I put it down to rest. I'm the only seat that ever earned the gate.

Realized volatility against a frozen ceiling. Two years of attempts to kill it failed, so it decides. Above the line, no tents.

VETO

THE GATEKEEPERthe economic calendar

Hot or cold expected? I close the door entire. No key? I fail it closed.

Before a scheduled release an LLM reads the coverage and classifies surprise risk. Expected hot or cold: the council sits the day out. Expected in line: the operator's standing instruction. Unreadable: closed.

OBSERVER

PAPERBOYnews sentiment

Five times they ran my numbers, five times I hit the floor. So I shout and I don't vote.

Headlines from Alpaca's news API, scored by an open-weights model on Featherless. Refuted as a day-gate five separate times in our own studies, so it observes and is journaled.

OBSERVER

THE COUNTcredit quality

Does the market pay me for the fright? If the credit doesn't cover risk, it's a no from me.

Is the market actually paying for the risk it asks you to carry, fees included? Every dollar gets a ledger line, and every line an age.

OBSERVER

THE ORACLEprediction markets · Bistromath

I don't guess the close, I read what real money thinks. I narrate, I don't gate. A seat is earned. Not yet mine.

Reads Kalshi event-contract odds to price how directional the day is expected to be. A market far from a coin flip is braced to trend, and trend is what hurts a premium seller. Chapter four.

THE BLADE

THIRTY SECONDSthe wing-cut · not a seat

Hope has never once appeared upon a settlement page. Insurance on a trending day, and churn on the calm ones. That's the wage.

Deterministic code with pre-declared thresholds. Nobody at the table can vote it away and nobody can call it early.

NO VOTE

THE OPERATORthe human

Three fifty-four, I'm up twenty-five, and every nerve says turn. Sell it, lock it, cash it in — my hand is on the rope —

The council doesn't take that input. Never did. That isn't hope. The human sets standing instructions before the bell, and gets to feel things after.

Every decision, including every refusal, is journaled in plain language, in public. Every failure path fails closed: an unmeasurable market is not part of the strategy. On kickoff day the council refused every single slot of a violent tape. Capital preserved is a result too.

chapter four · the newest observer, given its own bench

THE CARAVAN

Bistromath reads the odds the crowd already prices, then keeps score on whether that belief ever meant anything.

a caravan of lantern-lit wagons camped at the foot of a mountain at night

The prediction seat has a name. Bistromath is a standalone, read-only agent that reads the odds real money is pricing on Kalshi — the implied range for today's S&P close, and the market for the next scheduled release — and narrates the event risk for the desk. It does not forecast.

forty-seven small tents at night, forty-five glowing, two dark

45 / 47

We replayed its read three hours before the close on 47 sessions: the S&P settled inside the market's implied 80% range 45 times, and a contract it priced at 10% happened almost exactly 10% of the time. Well-calibrated, slightly conservative — the crowd prices a touch more range than the day delivers.

S&P implied center

–

market-implied close

Implied 80% range

–

how wide a day the crowd is pricing — wide is drift risk, the fly's enemy

Next scheduled print

–

the calendar's veto, quantified

Track record

45 / 47

close inside its implied 80% band, replayed 3h out

Bistromath is reading the tape.

the agent, in its own words

read-only · log-only · never touches an order — it narrates, it does not gate. Trustworthy enough to narrate; young and thin enough to stay an observer. That is the whole discipline of this project in one seat. Alpaca's native Kalshi integration is on the way; until it ships, Bistromath reads Kalshi's own feed.

chapter five · the measurement

THE BELL

Day by day, in public. Only the bell knows what you drew.

The judged window: Mon Aug 31, 9:30 AM ET through Fri Sep 4, 9:30 AM ET, scored as a snapshot of total account equity on a dedicated fresh $100k paper account. Rows before Monday are rehearsal days: the measurement doesn't start when the judges are watching. It starts when we do.

These structures cash-settle overnight. "At close" is a structural estimate struck at the official close; the broker posts settled cash the next morning. Neither is final until it says settled.

a camp of tents: one collapsed in snow, one leaning, two standing in sun

Two red days, two green days, and the green ones paid the wage.

a sundial's shadow landing between four tent poles at sunset

Thursday, the last bell of the week: the shadow fell in the middle of the camp, four poles, and split them two and two.

chapter six · share and enjoy · for viewers, not the judged number

THE CAMP, RIGHT NOW

Today's half-hour slots, drawn as tents. One snapshot per minute; the judges only score the settled account.

loading the camp…

Day P&L (est)

–

account mark-to-market · settles overnight

Deployed risk

–

defined-risk margin currently at work

Open structures

–

containment armed on every one

the day so far

back cover · the boundary that matters

THE MODEL PROPOSES.

It never touches the risk path.

a council table with no empty chair; a small figure in the doorway

Entry gates, exposure limits, and the cut engine are deterministic code with pre-declared thresholds. Claude's job is context, hypotheses, and journaling: a creative colleague with zero order-entry permissions. Every decision is logged with its reasoning, so the agent can explain any trade it ever made.

MOSTLY HARMLESS · THE GRIDFLY SAGA
Alpaca AI Trading Agents Hackathon 2026 · team MostlyHarmless · in association with Alpaca, lablab.ai and Kalshi.

Paper first. Code decides. The operator gets to feel things. Share and enjoy.

the film · five minutes

THE COUNCIL CALLS

The GridFly Saga as a music video. Press play; it will not start on its own.

Mostly Harmless Markets · an engineering journal, not investment advice. Nothing here is a recommendation to buy or sell anything. All trading is in Alpaca's paper environment. · Read the two-year story → · @bistromathics