The Trading Opportunity List Is Part of the Agent’s Strategy

46.5% of entries in our historical perpetuals fleet landed in symbols shown on its movers surface, against an 8.9% random-availability baseline. The opportunity list shaped what the agent traded. For an agentic trading researcher, that makes rendering an experimental variable rather than presentation polish.

The continuous record covers a June 8 to August 15, 2026 pre-alpha fleet, with 231,638 finalized turns and 14,596 fills across 500-599 agents over its history. Most fills came from a paper engine. These are historical research observations, with the paper’s fill assumptions, rather than a return assessment of today’s DXAP.

The discontinuity that matters

The descriptive share alone leaves an obvious alternative explanation: a movers list might expose assets that already deserve attention. We therefore need a comparison near the actual display boundary. The published regression discontinuity at rank three versus rank four estimated selection at 1.75 times higher, with an interval of [1.49, 2.06], at that boundary. Nearby symbols had statistically similar market state, while one side remained unshown.

That result supports a local causal interpretation about selection. It gives us a stronger engineering reason to inspect the display than the aggregate percentage provides. Its scope is the particular render boundary and setting. Moving it into a universal claim about every ranked table would lose the identifying comparison.

Three denominators for an audit

Our proposed audit separates symbols that were available to trade, symbols actually rendered to the model, and symbols subsequently selected. Each turn needs the rendered context as well as the chosen action. A market-universe snapshot alone cannot recover the second denominator.

The arithmetic illustrates why this matters. Dividing 46.5 by 8.9 gives approximately 5.22. That descriptive concentration ratio and the local 1.75 selection estimate answer different questions. The first compares aggregate exposure concentration against an availability baseline. The second compares nearby ranks around a rendering boundary. Treating them as interchangeable would exaggerate the causal estimate.

An evaluation should retain the original candidate order, timestamp and selection outcome. Replays can then change one rendering decision while holding the market snapshot and mandate fixed. We would register both selection share and economic outcomes before running that experiment, because increased selection frequency alone has no preferred sign for returns.

Turning exposure research into product criteria

The record places the value of rerouting at an honest $0-17K depending on assumptions. That range limits an economic claim even where the selection mechanism is better established. The engineering priority is to make candidate exposure inspectable and test its alternatives under declared assumptions.

Our controls paper supplies the broader trace discipline: preserve the path from mandate through rendered context to action. DXAP’s current public description includes a data index, model proposals, external policy checks and recorded turns. That architecture gives a reader concrete questions to ask about exposure logging; the description alone establishes no completed render experiment. Frontier progress here means measuring the choices the harness makes before the model chooses a trade.