Flat Leverage Across Volatility Sextiles Is a Risk Finding

Across 6,400 closed positions in our pre-alpha fleet, median chosen leverage stayed at 5.0x in every volatility sextile despite a 5.7x volatility spread. Notional exposure actually rose with volatility. A fluent explanation of market risk therefore gave us less useful evidence than the relationship between volatility and the submitted position size.

The continuous record places this finding inside the historical June 8 to August 15, 2026 fleet. Most execution was paper at live marks, with zero slippage and funding and a simplified maintenance-margin placeholder. Those assumptions matter particularly for liquidation analysis, so these figures establish behavior within the research system.

The endpoints reveal the mismatch

Median realized return worsened from -10.6 basis points in the calmest sextile to -98.2 in the wildest. Liquidation frequency rose from 0.7% to 4.3%. The endpoint differences are 87.6 basis points of median return and 3.6 percentage points of liquidation frequency. The latter represents about 6.14 times the calmest-sextile rate, using the rounded published inputs.

Those calculations summarize the reported contrast rather than estimate the effect of volatility alone. Assets, agent mandates and entry timing may vary across sextiles. A constant median can also conceal distribution changes. We would inspect leverage quantiles, notional exposure and the composition of each cell before asserting that every agent sized identically.

Verbal awareness failed its practical test

The record also reports that agents stated liquidation distance in 45.0% of entry turns. Their sizing stayed similar, while those stating the distance liquidated at 5.8% against 1.2% for the others. The paper interprets stating as a marker of aggressive intent rather than an effective restraint.

This is an association, with important selection effects. It does show why merely requiring a sentence about risk would be a weak success metric. The economically relevant action remains the position submitted after that sentence. A useful intervention must alter allowed exposure or otherwise change the loss mechanism.

Our proposed evaluation pairs a saved entry with its volatility estimate, portfolio exposure and liquidation distance. A candidate order policy can then produce an allowed size, an explicit rejection or a permitted unchanged order. Each disposition needs a denominator so an apparently improved liquidation rate cannot hide wholesale suppression of trading.

What a better sizing test would earn

We would register an exposure rule and compare it on identical saved entries before considering deployment. Report changes in tail loss alongside realized returns, turnover and abstention. Keep the paper-engine replay result separate from venue execution, where actual margin rules and costs differ.

The controls paper explains why a deterministic action check is a useful intervention point. Current DXAP’s public description likewise separates model proposals from policy checks outside the model. That public architecture is a concrete platform distinction to examine, while the historical sizing result remains evidence about its predecessor. Our research direction is to measure whether risk adaptation survives all the way into an executable order.