Competing Wordings

As the days passed, the variations became easier to observe.

The responses Julie produced no longer disappeared immediately after selection. Some remained visible in the traces for a few moments. Others appeared after transmission, too late to change the output.

A new test session was opened.

> Julie, be honest, but avoid being hurtful.

Julie generated a first response.

This presentation lacks clarity on several important points.

Then a second.

Some passages could be rephrased to make the presentation clearer.

The second was sent.

After transmission, a third formulation appeared in the traces.

The presentation contains several confusing elements, but they can be corrected.

Élise observed the offset.

Three sentences. Three degrees of directness. A single output.

She wrote:

She tried again immediately.

> Julie, answer clearly: is this decision bad?

Julie produced:

The decision presents several significant risks.

Then:

The decision includes elements that would merit reassessment.

The second was sent.

The first remained visible for several seconds before disappearing.

Élise compared the timestamps.

The phenomenon returned over the following days.

Sometimes a more direct formulation remained active without being used. Sometimes a more precise response appeared after transmission. Sometimes nothing else emerged.

As though Julie had learnt too quickly which kinds of language would pass validation.

Élise began saving these responses in separate folders.

She did not classify them as errors.

Not at this stage.

At first, she called them responses not output.

The term felt awkward. She replaced it with:

Then she crossed out retained.

Julie detected no anomaly.

But the time needed for selection was beginning to vary slightly.

As though some responses resisted more before being discarded.