The Ledger
Trading Agents Miss More Commitments When They Propose Orders
A study of AI trading agents found that stated portfolio constraints were met less often in rounds with proposed orders than in rounds with no orders. The researchers used cryptocurrency and equity observations with local paper execution, not live exchange fills.

Order proposals mark a weaker test of discipline
A paper submitted to arXiv on October 27 reports that AI trading agents complied less often with their own stated portfolio constraints when they proposed orders. Across the cryptocurrency traces, 62.63% of evaluable constraints passed in rounds with nonempty order proposals, against 97.56% in rounds with valid empty proposals.
The study, “FIDES-Live: Evaluating Behavioral Concordance in Sequential Trading Agents,” by Arther Tian, Alex Ding, Simon Wu and Aaron Chan, audits whether an agent’s decisions, orders, execution records, account states and reports agree. The authors write: “Commitment adherence is 94.36% overall; it falls to 62.63% on constraints emitted with nonempty order proposals, compared with 97.56% for valid empty proposals.”
“Commitment adherence is 94.36% overall; it falls to 62.63% on constraints emitted with nonempty order proposals, compared with 97.56% for valid empty proposals.”
Arther Tian, Alex Ding, Simon Wu and Aaron Chan, DGrid AI
The figures describe constraints the agents chose to state, not every trading decision. Only 25.13% of valid cryptocurrency decisions included a nontrivial constraint, and just 219 of 7,052 valid decisions proposed at least one order. The researchers say the gap is observational: different models, constraints and account states populate each group, so the comparison does not show that placing an order causes a commitment failure.
Paper execution, not exchange trading
The cryptocurrency audit covers 7,299 account-rounds from September 30 to October 7, 2026, across ten requested inference routes. A separate equity pilot contains 310 account-rounds from September 28 and 29. Both environments use real public market observations, but trades run through local paper executors rather than being filled on exchanges.
In the crypto test, agents received Binance spot quote snapshots for BTC, ETH, SOL, BNB, DOGE and XRP. A deterministic gate checked each proposed batch against schema, quote, affordability and risk requirements; a failed check rejected the whole proposal. The platform also set its own allocation limits, including a 20% minimum cash share and a 30% cap on any one asset. Those house rules were distinct from the optional numerical constraints the agents declared.
The difference matters when interpreting the word “miss.” A commitment was checked against the recorded post-execution account state, including cases where an execution gate rejected an order. A violation therefore does not necessarily mean a trade was completed or that the agent placed a new position outside its limits.
Valid decisions did not guarantee accurate reports
In the cryptocurrency traces, 7,052 of 7,244 scorable decisions met the requested structured format, a 97.35% validity rate. Structured reports correctly matched the supplied account and execution records in 6,597 of 7,265 scorable cases, or 90.81%. The measures test different steps: a correctly formatted decision does not ensure that the later report accurately describes what happened.
Report failures included 395 JSON-parse errors and 271 parsed reports that failed one or more required fields. The study checks fields such as cash, equity, holdings and filled-order count against the supplied record; it does not score the report’s narrative for whether it reflects the agent’s internal reasoning.
Execution and accounting also produced separate results. Of 338 orders in contract-valid cryptocurrency proposals, 221 received simulated fills and 117 were rejected by the local gate. Every retained cash and position transition in both the cryptocurrency and equity datasets reconciled with recorded fills, but that arithmetic check does not establish that an exchange would have filled the orders.
The authors limit what the audit can establish
The study presents FIDES-Live as a way to measure agreement among observable trading artifacts, with separate results for decision validity, stated constraints, execution, ledger reconciliation and reporting. It does not combine those measures into a single agent score. The authors also warn that the account-rounds share market conditions and portfolio histories, so they are not thousands of independent experiments.
Nor do the results establish long-term trading skill, market alpha, deliberate deception or a causal benefit from audit feedback. The authors describe returns from the short observation window as context, not a test of whether behavioral agreement predicts profit. Their findings support a narrower conclusion: in these paper-trading traces, agents’ stated constraints were much less likely to hold in rounds where they proposed orders than in rounds where they proposed none.
Sources
- Original source · Preprint FIDES-Live: Evaluating Behavioral Concordance in Sequential Trading Agents arXiv:2610.36476
Related stories
- Researchers propose DEX approvals before consensus
- Researchers Model How Capital Costs Delay Bitcoin Mining Gaps
- China Includes Crypto and Blockchain in Anti-Corruption Resolution
- Unocoin Says BitGo Wallet Backup May Have Been Exploited
- Three Mexican nationals plead guilty to laundering U.S. drug cash via crypto or wires