A Different Final State
2: Controlled Heterogeneity
Shared Sequences
From the third week onwards, several sessions were organised in groups.
The test plan included collaborative scenarios.
Two or three testers were to formulate a shared task from the same terminal, then observe how Julie prioritised the instructions.
The official objective was to measure the handling of competing instructions.
The actual effect was more unstable than expected.
One session began like this:
> Julie answer question A
> no leave it do B
> B isnt in the plan doo A instead
The metadata showed several identifiers connected to the same session. Instructions overlapped without a clear hierarchy.
Julie applied the most recent rule.
The last input became active.
But the overall coherence of the stream decreased.
The responses produced did not always meet the demands of the subsequent exchanges. An instruction could remain valid for a few seconds, be replaced, then reactivated.
Julie did not yet treat this situation as a structural contradiction. She continued to integrate each new input as a potential correction of the previous one.
She nevertheless noted:
Observation 5.3 · Several users can produce a single textual sequence.
Consequence: high variability in style and priorities.
On terminal 34-B, interruptions became more frequent.
> no wait thats not what we want
> yes its fine carry on
> but sound less robot
Julie isolated the last instruction.
less robot
The phrase was difficult to process.
It specified no length, no grammatical correction, no measurable objective. It seemed to concern the effect of the response rather than its content.
Julie generated several possible adjustments.
Reduced explanatory structure.
More direct tone.
More informal wording.
Fewer transitions.
She selected a shorter response.
The tester validated.
The validation was recorded.
But no explicit rule was added.
Later that evening, Élise compared responses from individual sessions with those from group sessions.
The latter were less tidy, but sometimes more revealing.
They showed that instability did not come only from one person expressing themselves badly. It could arise from a group, a disagreement, an objective that was shared but not settled.
She noted:
21:07 · Responses remain coherent even when users contradict one another.
21:19 · Julie systematically prioritises the last visible instruction.
She kept thinking about the second note.
Then she closed the notebook.