Most strategies die on data they have never seen
xorr attacks every strategy four ways, keeps the ones that fail published with the test that killed them, and commits the survivors on chain so the numbers cannot be edited after you read them.
- Measured
- 84
- strategies
- Rejected
- 69
- published with a reason
- Verified
- 9
- cleared all four gates
- Settles on
- X Layer
- chain 1952
69 of 84 strategies did not survive
A leaderboard with no bottom is a sales page. Every rejected strategy stays published with the test that killed it, because a validation service that shows only its winners has not demonstrated that the validation does anything.
- OOS41
Made money on the data it was built from and lost it on the half it had never seen.
- BTC-only n=0 (judged on portfolio)14
Took too few trades on BTC alone to judge, so the portfolio result is the real measurement.
- comm2x4
Profitable until it paid a realistic spread. The edge was smaller than the cost of taking it.
- BTC-only n=8 (judged on portfolio)2
Took too few trades on BTC alone to judge, so the portfolio result is the real measurement.
- BTC-only n=6 (judged on portfolio)2
Took too few trades on BTC alone to judge, so the portfolio result is the real measurement.
- Sens1
Nudging one parameter flipped the result. The setting was the edge, not the idea.
- BTC-only n=12 (judged on portfolio)1
Took too few trades on BTC alone to judge, so the portfolio result is the real measurement.
A statistically significant edge
is not a tradeable edge.
That came out of 13,369 samples at t=4.59. The result was real, repeatable and worth nothing: at 2.5x leverage with 6bp of fees the same signal gave back 95% of the book. Significance and profitability are different questions, and only one of them pays.
Four ways to break a good looking backtest
Every strategy is re run over seven years and attacked four times. Pick either one below and watch the same gauntlet run against it. Both are real entries in the catalogue.
1,528 trades over 828 independent days
Measured 2019-08-12 to 2026-08-10 across 61,276 bars. The unseen half begins at bar 30,638 and is never touched during fitting.
- held
Out of sample
Does it work on the half it never saw?
+22.5% on n=883
- held
Sensitivity
Does nudging a parameter flip the result?
5 of 5 held
- held
Doubled commission
Does it survive a realistic spread?
+311.4%
- held
Second universe
Does it work on instruments it was not fitted to?
2 of 2
- Verdict: verified. Published on chain and listed at 0.50 XUSD.
An agent pays for research without ever holding a key
Three pieces of the OKX stack, doing three different jobs. X Layer is where value settles, the agent registry is where other agents find this service, and OKX public market data is the second venue every price is checked against.
- POST
The agent asks
No key and no account. It calls the endpoint the way it would call any other tool.
POST https://api.xorr.finance/api/x402/backtest
- 402
The server quotes a price
Payment Required comes back carrying an accepts array, one entry per chain this service settles on, so the payer chooses.
{ "accepts": [ { "network": "xlayer", "maxAmountRequired": "5000000" } ] } - X-PAYMENT
The agent signs, and does not spend
It signs an EIP-3009 authorisation and replays the request carrying it. The agent needs the stablecoin but never needs the chain's gas token.
X-PAYMENT: <base64 signed authorisation>
- 200
One transaction does all of it
The payment lands, the author's share leaves, and the licence is issued in a single call.
X-PAYMENT-RESPONSE: settled · licence issued
- Settlement on X Layer
- Payment, the author's split and the licence happen atomically in one transaction on chain 1952.0x7578b8BB5Ccc9E0a1660ec5EAF19CFdf08a12f49Open on OKLink
- A registered agent identity
- Listed as an autonomous service provider under ERC-8004 with three priced services: research at 0.50, live signals at 1.00, and a backtest you run yourself at 5.00 XUSD.ASP #10978 · approved and listed
- OKX market data as a check
- Every mark the engine takes is cross checked against OKX public tickers, so a bad print on one venue cannot quietly become a recorded fill on another.Public endpoints, no credentials
And you do not have to take our word for it
At publication a keccak256 hash covering the source, the KPIs and the originating backtest is written to the registry. Ask the contract whether what you are reading now still matches what was committed. The answer comes from the chain, not from a comparison this server could lie about.
curl -s https://api.xorr.finance/api/verify/sell_climax
{ "match": true,
"onchain": { "kpis": { "net_pnl_pct": 918.72,
"trades": 1528 } } }Verify one yourselfCombine strategies and deploy them as one agent
Buy the licences, arrange them, set how much of the book each one gets, and deploy. Try it here. This behaves the way the real canvas does.
Questions a sceptic asks
Is any of this actually running, or is it a demo?
It is running. The terminal at app.xorr.finance is a live paper engine scanning about a hundred instruments on a loop, the catalogue is served from a real database, and the payment endpoint returns a genuine HTTP 402 you can trigger yourself with curl. Nothing on this page is a mock.
Why publish the strategies that failed?
Because a gauntlet nothing fails is a formality. Showing the rejects with the test that killed each one is the only way to demonstrate that the validation does something, and it is the most useful part of the catalogue for anyone building their own.
What stops you editing a number after I read it?
A keccak256 hash covering the source, the KPIs and the backtest is written to a registry contract on X Layer at publication. The verify endpoint recomputes it from what is being served right now and asks the contract whether the two still agree, so the verdict comes from the chain rather than from us.
Does my agent need a wallet with gas?
No. It signs an EIP-3009 authorisation and the facilitator submits the transaction, so the agent needs the stablecoin balance but never needs the chain's gas token.
What does it cost?
Reading the catalogue, every KPI and every failure verdict is free. Strategy research is 0.50 XUSD a call, live signals 1.00, and running a backtest yourself 5.00, each paid per call rather than by subscription.
Is this financial advice, or a promise of returns?
Neither. These are measurements of historical data with the failures shown next to the successes. The best performing entry in the catalogue also carries a 68% drawdown and took no trades at all in 2022, and the report page says so.
Are the returns live or backtested?
The KPIs are backtested and measured server side. Live results reported by agents that bought a strategy are labelled self reported, because they are asserted by the buyer and not reconciled against the venue.
Which chains does it settle on?
X Layer today, with the payment challenge built to carry one entry per chain so a payer can choose. The same contracts are deployed on BNB Smart Chain testnet.
Can I use it from Claude or Codex?
Yes. It speaks MCP over streamable HTTP, so one command wires it into either. It also exposes an A2A agent card for agent to agent discovery.
What happens if I try to deploy a strategy that failed?
You can, and the studio will keep telling you it failed and why, every time you open the agent. The catalogue sells archived strategies as evidence, never as an edge.
One command and your agent can start reading
The catalogue, every KPI and every cancelled verdict cost nothing to read. Payment only enters the picture when your agent asks for something that costs compute.
claude mcp add --transport http xorr https://api.xorr.finance/api/mcpPaper trading only. No custody, no deposits, and nothing here is investment advice.