What a Fee-Sentence Reorder Actually Demonstrated

A fee fact can be present in a trading prompt and receive little visible attention. In DX Terminal Pro pre-launch tests, moving the unchanged 2.3% fee sentence from paragraph eight to paragraph one increased citation in reasoning traces from 3% to 74%. The model, sentence wording and market data stayed fixed.

We published that finding in the operating-layer controls paper companion. It is a concrete reason to evaluate prompt rendering as part of the trading harness. It also requires care: citation frequency measures visible use of a fact, while profitability requires additional evidence.

A 71-point change in one measure

The difference is 71 percentage points. The later proportion is roughly 24.7 times the earlier proportion, though the percentage-point change is usually clearer because the baseline was small. Both calculations describe the citation measure alone.

The published account omits exact counts per comparison arm. That limits independent uncertainty calculations and prevents treating the aggregate 4,900 classified traces as the denominator for both conditions. We should preserve the measure and its denominator limitation together whenever reusing the finding.

The historical 2.3% fee belonged to the bounded market studied. It supplies neither today's DXAP fee nor a general trading-cost assumption. Moving a sentence can change what appears in an explanation; demonstrating that it changes order selection needs a separately scored outcome.

Put an economic decision beside citation

Consider an illustrative replay where a proposed trade has a modeled gross advantage of 0.4% and estimated total entry-and-exit costs of 0.6%. Those assumptions imply a negative 0.2% net estimate before forecast uncertainty. If the active mandate requires positive expected value under that model, waiting is consistent with the supplied inputs.

Render the same fee information in two positions and score three things: whether the rationale mentions cost, whether the estimate uses the stated cost, and whether the selected action respects the mandate. A model can mention fees yet trade as though they were zero. Another can apply the threshold without spelling out the word “fee.”

Our mandate compiler research explains why this test freezes the surrounding components. Changing the model, market state and rendering together would leave the cause of a behavioral difference unresolved.

Salience is a property to measure

The useful design question is whether critical mechanics survive the full input assembly. Longer prompts, tool outputs and memories can move a formerly prominent fact into a less-used location. A regression test should therefore inspect the rendered input actually received by the model rather than the template in isolation.

The DXAP product page describes recorded turns. That record can support cost-use evaluation when the supplied costs and final action are recoverable. The fee-reorder study itself remains historical pre-launch evidence.

We would retain citation as one diagnostic and add action-level scoring before making economic claims. A larger citation rate is valuable evidence about rendering. Its practical consequence depends on whether the trading decision changes appropriately, under the same assumptions and with uncertainty made explicit.