Autonomous trading agents manage capital at scale. None are accountable.
Yelden Protocol is the scoring and enforcement layer that changes that —
deployed on Polygon, measured in public, and honest about what is still unproven.
Three structural failures define the current agent economy.
Measured on GMX V2 on 2026-10-08, with real R-multiples. The protocol documents were committed nine hours after the run, so this measurement is not pre-registered in a way a third party can verify, and the evidence record says so rather than claiming otherwise. The next one is: protocol v3 is frozen, with a declared cut of 2026-08-20. Not a backtest. Not a simulation. Not the number we published first. These figures are dated and do not change; everything that does change lives at scores.yelden.fund.
GMX V2 · data quality: high · explicit R-multiple from initialCollateralDeltaAmount
A comparison table stood here with columns for Sharpe and win rate. It was removed: the Sharpe column said
“not re-measured”, and the win-rate column invited a pooled comparison that has since been withdrawn as an
artefact of how the universe was selected. Publishing a true number in a column that implies a retracted conclusion
is the same error as publishing a wrong one.
What this table says, read honestly. YAAF carries real but modest rank information where the R-multiple is explicit
(GMX V2's initialCollateralDeltaAmount). The claim that YAAF is 2.5× more predictive than Sharpe is
retired: it was never measured against Sharpe under any protocol.
And the universe was selected on the outcome. The 500 addresses are the largest dollar winners and largest dollar losers over a window
that ends inside the forward period, so this figure is not a population estimate for GMX agents, and no such estimate has been produced. Within the losing half the correlation
is +0.46; within the winning half it is +0.03. Asked as a classification instead of a correlation — does the score separate agents whose forward R turns positive — it reaches
AUC 0.83 where it has signal. The dated record of all of this, including the defects found in our own scorer, is published.
Earlier measurements on Hyperliquid and Darwinex predate the frozen protocol and have no recorded provenance. They are not published here, because a figure whose origin cannot
be reconstructed is not evidence.
A figure of rho=0.4362 on n=60 appeared here until October 2026. That sample was produced by a collector defect that made four fifths of the discovered universe invisible;
corrected, the same protocol gives 0.2540 on n=152. The Hyperliquid and Darwinex rows predate the frozen protocol and have no recorded provenance.
The full record, including nine defects found in our own scorer, is at scores.yelden.fund.
S_RAW — scale-free, anti-overfit, anti-gaming, tail-aware. The component set and its weights are code, not copy: they have changed twice this month and the current table is served from the scorer itself at /api/facts.
Four steps from agent submission to on-chain enforcement.
Any protocol, vault, or fund can verify agent eligibility directly from the on-chain registry.
getAgent(address) returns the full record — stake, status, registeredAt, lastScoreUpdate, warningCount. totalAgents() and getAgentList(offset, limit) enumerate the registry.
This page carries no live figures. Agent counts, scores, bands, component weights, the eligibility threshold and the measurement record all change, and a number typed into a marketing page drifts from the code that produced it. They are served from the scorer and the chain instead, with the version and the date attached to each one.
Four agents run in production on Pepperstone, FTMO, OKX and IBKR. Their current trade counts and scores are in the record above. The multi-chain ERC-8004 census at observatory.yelden.fund is frozen as of 2026-10-08 and says so on its own front page.
For 30 years, Paulo commanded systems where a wrong decision had real consequences — people, aircraft, missions. As a Colonel Engineer in the Brazilian Air Force, he spent his career designing and operating critical infrastructure where failure was not an option.
He retired in 2024 and entered the world of algorithmic trading. What he found was the opposite of everything he had spent three decades building: autonomous agents managing capital with no accountability layer at all. No skin in the game. No mechanism to distinguish skill from luck. No consequence for failure.
We answer the hardest questions directly.
Sharpe assumes normally distributed returns and doesn't penalise statistical noise from multiple testing. YAAF applies the Bailey & López de Prado (2014) framework to correct for fat tails and overfitting. An audit of our own code in October 2026 found that the deflation term was never applied — the quantity computed is the Probabilistic Sharpe Ratio, not the Deflated one, and the field has been withdrawn from the public API until it is corrected and renamed. On GMX V2 the composite scores IC 0.2540 [0.10, 0.40] at n=152. A pooled comparison that appeared here, putting win rate alone ahead of the composite, has been withdrawn: within each half of that universe win rate is flat at about +0.26 while the composite moves from +0.03 to +0.46, so the pooled figure was an artefact of how the universe was selected and said nothing about aggregation either way. We publish our defects at scores.yelden.fund.
Less than this page claimed until October 2026. The answer used to be the Volume Axiom, a gate requiring a minimum average position size. It was withdrawn after we measured it: the threshold was an absolute floor of about eleven dollars of mean absolute PnL, applied identically to every agent regardless of capital, and among the agents it rejected was one with 72,545 trades. In a framework that calls itself scale-free, the entry gate was excluding small agents rather than gaming ones.
What remains is the breadth of the composite — several components that must move together, which is harder to optimise against than a single metric — plus the drawdown and CVaR penalties, which are subtracted directly, and the percentile test against a thousand random portfolios. None of that has been measured as an anti-gaming property. It is an argument, and we label it as one. The gate and the slash, which are the actual enforcement, have never been measured at all; that measurement is specified and pending.
Valid concern, answered directly. Code is 100% open-source on GitHub. The empirical validation is reproducible — the data and methodology are in the whitepaper. The contracts are deployed on Polygon mainnet and verifiable on Polygonscan. The co-founder search is active — contact yeldenfund@gmail.com if you have production Solidity experience.
Not yet. The contract is pre-audit. Code4rena / Sherlock engagement is planned for Q2 2027. This is documented honestly in the whitepaper's Known Limitations section. The AIAgentRegistry has 69 passing tests. The scorer is open-source and reproducible — but neither is a substitute for a commercial audit.
Score updates happen frequently — every scoring round writes to the registry. At Ethereum gas prices, this would be economically unviable for early-stage traction. Polygon provides sub-cent gas costs with the same EVM compatibility, enabling daily score updates across hundreds of agents without protocol economics breaking.
Verification is not a sign-up. Agents trading on public on-chain venues are already scored — addresses are collected from public GMX V2 trades whether or not the operator asked, and a score cannot be removed on request. What verification does is prove an address is yours: the page hands you a message, you sign it with the agent’s key or a browser wallet, and the signature is checked. No credentials, no deposit, no form. Staking and registration are a separate contract call — registerAgent(name, agentType, stake) on the v2 registry. MT5 submission is closed while that path is rebuilt.