When Chosen Leverage Behaves Like a Configuration Constant

Chosen leverage in our historical live fleet behaved substantially like a configuration constant. Risk tolerance mapped to leverage at +0.425x per slider level, agent fixed effects absorbed 60% of its variance, and realized leverage tracked explicit numbers in strategy text at Spearman 0.836. Those findings change where we would look for a risk-control improvement.

They come from the pre-alpha record summarized in our continuous-record companion. The fleet ran from June 8 to the August 15, 2026 cutoff, with 500-599 agents across its history. Its predominantly paper execution and changing model mix limit generalization to live venue outcomes and to current DXAP.

Configuration fidelity answers one question

Following a user’s declared number can be evidence of instruction fidelity. It becomes a risk problem when the same number survives materially different market conditions without an appropriate external constraint. A trader might have intended a maximum of five times leverage rather than a standing instruction to use five times leverage whenever entering.

We would therefore separate three objects in an audit: the owner’s authorized ceiling, the model’s proposed leverage, and the value allowed by the order policy. They can all equal five while expressing different responsibilities. The first records permission; the second records behavior; the third records enforcement.

Spearman correlation measures rank association. The published 0.836 supports strong tracking of user numbers, while leaving the cause of profitable or unprofitable outcomes unresolved. Likewise, the fixed-effects result describes variance explained by agent identity. It supports persistence across agents’ choices rather than an estimate of how much safer any specific policy would become.

A response curve instead of a reasoning quote

An illustrative audit can group identical mandates across several volatility ranges and plot proposed leverage against each range. The expected response must be specified by the mandate and risk policy before judging the model. A user explicitly requesting constant leverage and a user requesting constant risk warrant different reference curves.

Across four slider increments, the reported +0.425x association corresponds to 1.70x in the fitted relationship. This arithmetic explains the size of the configuration gradient; it is neither a guaranteed per-agent change nor a causal slider intervention. Retaining that distinction prevents a descriptive coefficient from becoming a product promise.

Our next proposed comparison would replay the same saved entries with an explicit volatility-sensitive policy. It would measure rejected orders, allowed exposure and simulated loss together. Counting lower proposed leverage alone could hide a rise in abstention or a change in the assets selected.

Put the check where exposure becomes real

The first controls paper records the importance of binding policy checks to exact typed actions. That is the useful connection: the risky parameter exists in the submitted order, so a check should inspect that order.

DXAP’s public homepage describes a model proposal followed by policy checks outside the model. We regard that separation as a meaningful criterion for comparing agent platforms. A current implementation audit and matched economic evaluation would establish how a particular configuration behaves today. The historical finding motivates that work; it does not certify its outcome.