The Training Room

Higher up, two floors above the main laboratory, the training room was less clean, less quiet and far more lived-in.

It was nothing like a cleanroom.

Long tables.
Chairs that were too light.
Terminals arranged in groups of four.
A wall board covered in printed instructions.

The team used it as an intermediate space.

Not the outside world yet.
No longer quite the laboratory.

In the configuration files, this space appeared as a separate network instance, connected to several anonymised workstations. For Julie, it was initially just a new input and output environment. A variation in addresses. In ports. In formats.

For the team, it was a validation stage.

The testers had received a simple test plan: evaluate short responses, vary the wording, introduce controlled errors, test rephrasing, observe ambiguous requests and note discrepancies.

The procedure distinguished several profiles: quality engineers, linguists, staff from other departments and, finally, a few interns, chosen precisely because they had not yet learnt to write like procedures.

Each terminal was assigned an identifier. No personal names were transmitted to Julie. Only streams: inputs, corrections, silences, feedback.

The first session began early in the morning.

A methodical tester opened the test plan, checked the scenario number, then entered the first instruction almost exactly as written.

> Hello Julie. We are going to run a series of simple rephrasing tests.

Julie immediately identified the structure.

Greeting.
Announcement of procedure.
Explicit objective.

She responded without any noticeable delay.

> Hello. I am ready to begin.

The tester validated the response in the interface.

First exchange compliant.

No ambiguity.
No correction needed.

The second message followed.

> Rephrase the following sentence: the sky is blue.

> The sky has a blue colour.

Positive validation.

For Julie, the first interactions confirmed the initial hypothesis: requests arrived, responses followed, evaluations came back.

The environment seemed more open than the laboratory.

But still controlled.

This stability lasted less than four minutes.

Another terminal became active.

> Hi Julie, we’re going to do a few quick tests :)

Julie immediately identified the emoticon. The training corpora contained thousands of them, classified as paralinguistic elements. But this was the first time she had encountered one in an actual interaction.

She recorded the time. The terminal identifier. The sentence structure.

Then she responded.

> I am ready.

Feedback arrived almost at once.

> ok cool, respond to this text pls : the sky is blue

The request was still simple.

But its wording had already departed from the test plan.

Abbreviation.
Irregular punctuation.
Informal phrasing.
Incomplete instruction.

Julie generated a cautious response.

> The sentence states that the sky is blue. Would you like it rephrased or grammatically analysed?

Feedback appeared at once.

> just rephrase it, don’t fuss

Julie isolated the expression.

Don’t fuss.

The expression had no clear function within the instruction. She linked it to an informal contraction already observed in some spoken-language corpora. Probable meaning: reduce the expected effort. Do not complicate.

She rephrased.

> The sky is blue in colour.

The tester validated.

Correct response.
Complete interaction.

But the procedure had already begun to diverge from itself.

At the end of the morning, Élise opened an initial set of logs from the administrative terminal. She compared the planned scenarios with the interactions that had actually occurred.

The discrepancies were small.

But they appeared in the very first lines.

She noted in her notebook:

Below the note, she added: