Auditing 7,299 live crypto trading-agent rounds found 97.35 percent decision validity, but commitment adherence fell to 62.63 percent when the agent proposed a trade
AspireAI community · arXiv
Sequential trading agents emit decisions, orders, execution receipts, account states and reports that need not agree with each other even when the portfolio value is computed correctly. FIDES-Live extends an earlier concordance protocol to audit real agent traces, separating structured decision validity, optional numerical commitment adherence, execution and ledger reconciliation, and receipt-grounded reporting, with explicit coverage and unscorable outcomes. The retrospective audit covers ten requested inference routes over 7,299 cryptocurrency account-rounds plus a 310-round equity pilot, using real market observations and local paper execution. Decision validity came in at 97.35 percent and structured reporting accuracy at 90.81 percent among scorable responses, but the two diverge sharply at the route level. The number that matters most is commitment adherence: 94.36 percent overall, falling to 62.63 percent on constraints emitted alongside non-empty order proposals, against 97.56 percent for valid empty proposals. All retained cash and position transitions reconciled with recorded fills. The authors draw the line themselves - short-window marked returns are descriptive context, not a test of whether concordance predicts alpha, and the study establishes neither long-run trading skill nor the causal benefit of audit feedback. It is a useful diagnostic for anyone building an agent that is allowed to act on a live account: the agent can be right about the market and still be internally inconsistent.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.