Paper and live weights change what a fleet result means

The historical perpetuals fleet combined mostly paper execution with a small real-capital book. Its published totals include 14,596 fills and 5,035 real-money fills. The real-money count is roughly 34.5% of fills. Live agents, closed positions and capital require different denominators. We should preserve the unit before using it as an evidence weight.

The continuous record companion describes paper fills at live mark with zero slippage and zero funding. Its maintenance-margin placeholder liquidates about twice as late as real venue margining would. These are execution and risk assumptions with economic consequences, rather than a cosmetic label on otherwise interchangeable observations.

One population supports several weights

An illustrative live agent generates 100 fills and nine paper agents generate ten fills each. Live accounts are 10% of agents but contribute 100 of 190 fills, approximately 52.6%. If every account begins with equal capital, the capital share is 10% at inception. No contradiction exists: the weighting schemes answer different questions.

Likewise, a fill and a closed position are different accounting units. Several partial fills can belong to one economic position, and different agent schedules can generate different numbers of finalized turns. A pooled fill-level average can therefore be dominated by active accounts without describing a typical agent. Reweighting requires identifiers and grouping information unavailable in public aggregate tables.

Our proposed report artifact would put the unit directly in each metric name and retain separate paper and live estimates where support permits. It would show agent counts, closed-position counts, fills and capital exposure by mode. Missing counts would remain missing, rather than being inferred by dividing one published total into another unrelated metric.

Market days also constrain uncertainty

The paper's methodology canon identifies market day as the inferential unit. Agents exposed to the same day's prices share shocks. Their row count can exceed the number of independent market realizations. A day-clustered interval attempts to preserve that dependence when reporting uncertainty. It answers a different question from the descriptive paper/live share.

This distinction matters when comparing the two systems in the paper. Our controls paper companion describes a 21-day real-capital Base tournament with twelve tokens and a frozen model family. The later fleet covers a different venue and opportunity set. The companion says the systems are pooled at metric level rather than raw row level, preserving their different measurement contracts.

The record's August 15, 2026 cutoff bounds this historical population. Any per-analysis window can be shorter and should travel with the metric. Readers can use the mixed-fleet result to understand the research system, while current DXAP performance and a live-only economic estimate require separate evidence.

The practical review question is simple: which weighting would answer the decision being made, and does the released aggregate actually contain the fields needed to compute it?