Normalize Fee Eras Before Comparing Agent Returns

Our historical fleet crossed builder-fee eras of 4.5, 14.5 and 5.5 basis points per side. Comparing raw returns across those eras would mix strategy behavior with a changed cost schedule. The paper therefore restates cross-era P&L at a common 5.5 basis points per side, giving the comparison a consistent accounting assumption.

The continuous-record companion describes the June 8 to August 15, 2026 pre-alpha fleet, with most fills coming from a paper engine. Its published aggregate P&L comparison covers June 8 to July 26. The date difference matters: the full data cutoff and a particular return window are separate boundaries.

A common rate solves an accounting problem

For an illustrative $10,000 fill, 4.5 basis points costs $4.50, 14.5 costs $14.50 and 5.5 costs $5.50 for that side. A same-notional entry and exit at 5.5 basis points would total $11 for those two builder-fee charges. This arithmetic deliberately excludes venue charges, funding, slippage and other costs.

Restating every fill at one rate holds the accounting rule constant. It does not recreate how agents might have behaved if they had faced that rate at decision time. A price-sensitive model might trade less, choose a different asset or hold longer when shown a higher fee. An accounting normalization preserves observed actions; a behavioral counterfactual needs its own experiment.

Our proposed comparison worksheet keeps three outputs visible: original-era accounting, common-rate restatement and zero-fee sensitivity. Each output uses the same declared trading population and window. It also identifies which cost components are being changed, because removing one fee component is different from eliminating every implementation cost.

The zero-fee result constrains the explanation

The record reports -$217K cumulative realized P&L at the common rate and -$148K at zero fee over its stated return window. The difference is $69K under those published accounting assumptions. The remaining negative outcome shows that fee removal alone fails to turn this historical fleet profitable.

That is a stronger decision implication than simply saying fees matter. It directs the next experiment toward entry quality, loss tails or exit behavior while keeping costs explicit. The paper’s directional-edge null and bracket replay help separate those research paths.

The first controls paper supplies a different cost environment: Terminal Pro charged 2.3% per swap in a bounded Base market. Converting that charge to 230 basis points explains why raw cross-system P&L would be a poor measure of model progress. Venue design and action opportunities changed alongside costs.

Current pricing requires current evidence

DXAP’s current homepage described no subscription and a 0.025% charge on agent volume when inspected on October 3, 2026. That is a public product description, rather than an all-in execution quote. We would verify venue fees and other costs separately for a current comparison. An evolving platform earns a useful economic assessment through transparent assumptions and matched outcomes, rather than by comparing today’s price with an unnormalized historical return.