A candidate list changes what an agent strategy means

A trading strategy includes the process that decides what the agent gets to inspect. Our historical rendered-candidate finding makes that claim measurable: selection changed at a display boundary, even among nearby ranks with comparable market state. For researchers, this means that a strategy inferred from chosen trades mixes the model's preference with the opportunity exposure supplied by the harness.

The continuous-record companion describes a historical Hyperliquid research fleet whose movers list displayed nine symbols: three gainers, three losers and three volume leaders. Rendered symbols received 46.5% of entries against an 8.9% random-availability baseline. That contrast describes concentration. The stronger identification comes from the rank boundary.

Section 4.2 of the primary paper reports a regression-discontinuity selection ratio of 1.75x, with interval [1.49, 2.06], at the gainers rank-3/rank-4 cut. Figure 3 covers 2,339 entry observations across gainers ranks 1 through 15 from June 16 to July 24, 2026. The raw rank-three/rank-four count ratio is separately 290/160, or 1.81x. Keeping those estimates distinct preserves the published analysis.

Two models of the same trading history

Imagine interpreting a fleet's momentum-heavy entries through a reasoning-only model. The agent reads market information, prefers rising assets and expresses that preference in its actions. The resulting trade distribution looks like evidence for an intrinsic momentum strategy.

Now introduce an exposure model. The renderer systematically highlights recent gainers. The agent chooses partly from the highlighted opportunities because they are available in the immediate context. The same trade distribution can arise with a weaker intrinsic preference. A selection jump at a display cut helps separate these accounts because the exposure changes sharply while the nearby ranking information changes gradually, under the discontinuity design's assumptions.

That local result has a precise reach. It supports an effect of rendering at the studied cut. It supplies neither an estimate for every unseen instrument nor a universal preference effect across arbitrary list designs. Comparability around the boundary and the absence of another discontinuous change remain part of the causal interpretation. The 46.5% concentration statistic alone cannot supply those conditions.

This distinction changes how we identify a strategy. Observed choice is conditional on an exposure process. A model that consistently selects from one rendered list may behave differently when another renderer supplies a broader or differently ranked universe. Describing the first behavior as the model's stable investment philosophy would hide a material harness assumption.

What the historical record changed

The continuous record expanded our earlier controls research into a fleet-level diagnosis of rendering and order-path mechanics. Its proposed response includes deliberate A/B evaluation of the render. The record establishes the historical finding and the research direction; it supplies no receipt that a particular replacement renderer shipped or improved current DXAP returns.

The economic distinction is equally useful. Changing exposure can reroute entries while worsening execution costs or concentrating another risk. The paper's rerouting economics depend on assumptions, with a stated $0 to $17K cost range. Selection sensitivity therefore deserves its own result before anyone calls it a profit opportunity. The historical fleets also showed no directional edge.

We would ask builders to state the renderer whenever they describe a model's strategy: what was eligible, what was exposed and how the list was formed. This is more than a disclosure detail. It identifies which part of the decision system a claimed preference belongs to. A strategy that survives an exposure change becomes a stronger model-behavior finding; a strategy that follows the list becomes a harness finding worth engineering directly.