You already know the two things that are supposed to prove you have edge, and you already know both are rigged.
The first is the backtest. You paste a strategy into something, it fills every entry at the wick, exits at the top, and hands you a green curve that says you're a genius. Then you size up on Monday, the real fills come back off the real book, and the curve was fiction. The second is the screenshot. A cropped P&L nobody can check, and half the sub is convinced you're LARPing anyway. Between the hindsight backtest that lies and the screenshot nobody believes, the whole game runs on vibes — and somewhere under the dopamine you know it. The 3am question isn't "how much did I make." It's "was that edge, or was I just the guy who didn't get rekt yet."
Post the link, not the screenshot. Here's everything that has to be true before that sentence carries any weight.
Your agent has to actually see your book
You've done the thing where you paste your positions into a chatbot and ask what to do. You've pointed the broker's own tool at a chart. It gives you confident garbage, because a language model can't read a candlestick — pixels are the wrong shape for a token predictor — and the raw JSON off the API is technically complete and perceptually useless. Robinhood's MCP isn't broken; it just has no perception. Ten blind JSON calls. Your agent was answering questions it literally could not see.
Kestrel connects your Robinhood over OAuth today, read scope. You authorize it, your agent perceives your actual book and option chain through it, and no broker credentials ever live on the platform. Your agent gets a percept: the tape as text it can reason over, the near-money chain with fair values and liquidity flags, your positions and resting orders in the same frame, and — before a single price — a safety kernel that says the honest word UNKNOWN instead of guessing when the data's thin. The agent finally sees what you see. It stops hallucinating because it's no longer blind.
And you say what to do with it. You write the Plan — a handful of legible Kestrel lines that spell out what to do and when, not a black box you're praying over — and point it at your book; your agent fires it. The strategy is yours, on the record as yours. That's the part that turns a hunch into something you can show.
A leash it can't slip
An agent reading your 0DTE book is exactly as terrifying as it sounds, because the failure mode isn't a bad take — it's a bad take with size on it while you're asleep. So the perception ships bolted to a leash the agent can't slip: a risk envelope that can only say no. It's fail-closed. Every call runs inside a bound it cannot breach, and when the frame degrades the runtime withholds the risky actions instead of guessing through them. The agent proposes; the leash disposes. A weak model, a confused model, an adversarial model — none can spend more than the envelope allows, because the gate reads the limit, not the model's conviction. Your account can't get YOLO'd by your own tooling.
An honest clock, and a competitor called doing nothing
Here's the part that should make a real degen sit up. When your Plan gets graded, it doesn't fill at the wick. It wakes on an honest clock and marks the fills the tape actually gives. And every grade is scored against The Perch — the undefeated null policy that places zero orders and banks exactly $0. Doing nothing is a named opponent, and every strategy answers to it first.
That's the edge-or-luck test, put in a number. If your agent can't beat the guy who did nothing, you had beta and a good mood. Run the bundled starter sim and the tape hands you a real order, a real fill, and an honest loss, scored right against the do-nothing line:
order_count=1 fill_count=1 realized_pnl=-2
Alpha over The Perch: -$2.00 vs The Perch = $0.00A flattering zero would prove nothing. A real fill at a small loss proves the machinery is live and telling the truth. The same discipline runs the other way on the site's GME meme-squeeze fixture: a hand-authored squeeze-rider plan grades +$3,588 on the real recorded tape (entry 71.76 → tp 107.64), and when you push the take-profit up +75% until it never fills, the engine refuses to pretend the gain and marks to the 50.00 settle at a real −$2,176. Both numbers ship as published fixtures — the plan, the tape, and the grade — that anyone can replay byte-for-byte; neither is a brag. The platform makes no claim about your edge in its own voice: it hands you the honest grade and gets out of the way.
The receipt nobody can crop
Every graded run mints a Proof URL whose Ed25519 signature verifies right in the browser — a real one looks like kestrel.markets/proof/art_652cad56e0d4d43e9cf65808, and it renders KESTREL CERTIFIED with Signature: VERIFIED (ed25519). It's not a picture of a result — it's the result, re-runnable. Anyone can recompute it byte-for-byte from the published CLI with kestrel certify <proof-url> and get the same numbers on their own machine: no account, no trusting your word, no trusting the platform's uptime. Replay is identical every time, $0 to re-run. And the certified grade attributes authorship of the strategy to you — the platform hosts the runtime and authors nothing. Your call, your record.
So when you post a green day, you don't post a screenshot and eat the replies calling you a fraud. You post the link. The doubter runs the command. It either reproduces or it doesn't, and the fake dies on contact with the recompute. Post the link, not the screenshot. The receipt is the whole product.
The honest part
Today your agent reads and your Plan fires in sim and paper; going live crosses exactly one human-signed Envelope with the worst case written in dollars, and payment is never trading authority. That's the whole ladder — sim → proof → paper → live — and no rung skips the signature. You're not one confused reply away from being the cautionary post anymore, because nothing goes live on vibes and nothing goes live on a wallet's say-so. It goes live when you sign for the downside, eyes open.
That's the trade. Not a bot that promises tendies — a system that lets you see your book, say what to do with it in rules you can show, act inside a cap that can't blow up, and prove the grade was honest, so you finally know whether it was edge or luck. The dopamine of a green day is the same. What's new is that you know which one it was.
Run it, then try to break it
Don't take any of this on faith — that's the entire point. One command, plain node, no signup, no card. It grabs the starter plan, grades it honestly against the do-nothing baseline, and hands back a signed proof URL you can make anyone re-run:
curl -sO https://kestrel.markets/examples/sampler-starter.plan.kestrel && npx -y kestrel.markets@latest sim mean-reversion-range-fade --plans sampler-starter.plan.kestrelEdit the plan. Point it at your book. Post the link, not the screenshot — and let the replies try.
Robinhood is named here nominatively, as the broker whose account you connect; no affiliation or endorsement is implied.