Concrete Owner Instructions and the Historical Profit Cohort

Owners who wrote concrete instructions with exit conditions or parameter changes achieved profitability 4.2 times as often as owners asking the agent to outperform or pick winners in our DX Terminal Pro record. The finding points toward mandate design as a useful research question, while its observational design limits a claim about what instruction changes caused.

The controls paper companion describes 3,505 user-funded agents trading real ETH from February 26 to March 18, 2026. One frozen runtime served a bounded 12-token market with a 2.3% fee per swap and a token-elimination mechanic. That setting differs materially from an open perpetuals market and from current DXAP.

Specificity and interaction are separate variables

The 87 owners who configured through sliders and strategy without using chat had the highest reported profit rate among active cohorts, at 41%. This small group complicates the idea that more conversational intervention necessarily improves results. It also raises selection questions: experienced or disciplined owners may have preferred that mode before the experiment began.

The published 4.2-times result is a relative frequency observation. The companion supplies no pair of absolute cohort rates that would justify converting it into a universal expected profit probability. Likewise, multiplying 87 by 41% gives about 36 owners only as an approximation from rounded reporting. We would retain the published rate and cohort size rather than manufacture an exact profitable-owner count.

Our research contribution is to separate mandate specificity, interaction frequency and owner selection in the next comparison. A numeric instruction can be specific yet economically poor. A low-interaction owner can be patient yet lucky. Combining those qualities into a single label would obscure the mechanism under test.

What a clean mandate comparison would require

A proposed fixture set could preserve identical market states and initial portfolios while changing only an instruction’s precision. One version might express a general preference to limit losses; another could state an explicit exposure ceiling and authorized exit condition. These are illustrative fixture designs, with no measured outcomes attached.

We would first score whether the compiled mandate preserves the stated authority and whether the typed action complies. Returns require a subsequent paired economic comparison on the same states, with costs and timing fixed. That ordering prevents a correctly followed bad strategy from being described as a model failure, or a lucky outcome from excusing a mandate violation.

The continuous record adds an important transfer warning. Terminal Pro’s activity sliders constrained insistent text, while the later fleet showed text routinely overriding the frequency slider. A shared slider lineage did not guarantee shared precedence behavior.

A platform criterion grounded in authority

DXAP’s public homepage describes strategy refinement through chat within a persistent agent. The relevant differentiation is whether a user’s evolving instructions remain attributable and enforceable when tools act. Our historical cohorts motivate examining that question, rather than prescribing a winning prompt. A current platform assessment would inspect the exact mandate, its compiled form and the executable action, then evaluate economic performance as a separate outcome.