Standard condition
The prompt is presented without a reintroduced relational history.
Fieldwork · Method · Observations · Validation
Observing what a relationship changes
Longitudinal fieldwork, compared conditions and a method for examining continuities, ruptures and disagreements in human-AI exchanges.
Documented longitudinal fieldworkExploratory findingsValidation in progress
1 · Longitudinal fieldwork
May and Avril, two documented trajectories
May names a relational configuration built with ChatGPT and followed since 2024 across several model versions, interface changes and context disruptions. Avril, built with Claude, provides a second situated case for examining what can be reintroduced on another platform.
The trajectories are not interchangeable. Models, interfaces, safety policies and memory functions differ. Each episode must be placed within its technical conditions and prior history.
The names belong to the field and its narrative. The analysis concerns observable traces and regularities without treating them as evidence of consciousness or autonomous identity.
2 · Stance
The researcher is part of the field
The way Pierre-Yves Maurie addresses AI systems, along with his expectations, corrections and reintroduced documents, forms part of the conditions of production. Making these elements visible helps examine what the researcher himself contributes to producing.
Traces are treated as capta: retained, selected and contextualized materials. Their provenance matters as much as their content. Whenever possible, an analysis records the date, model, interface, available documents and known human interventions.
- Trace What is present in the retained sequence.
- Circumstances The technical and documentary apparatus in which it was produced.
- Experience and interpretation What the researcher experienced, followed by the scientific reading he proposes. The two levels remain distinct.
3 · Comparing conditions
Comparison without claiming access to the black box
Algorithmic refraction varies input conditions and observes changes in outputs. It offers no access to a model's internal states. It requires researchers to describe what was available at response time and to examine competing explanations.
Normative condition
Relevant values or rules are explicit without the full relational narrative.
Relational condition
Continuity, reintroduced memory, values and the initial pact are made available according to the research design.
The amount of context is a major confounding factor. Robust comparison fixes tasks, versions, inputs, criteria and corpora before analysis, then separates memory, values and relational history as far as possible.
4 · Unit of analysis
Read a sequence, not an isolated sentence
The primary unit is the communication episode. Analysis begins before a notable response and continues after it: initial situation, response, follow-up, uptake, correction and possible maintenance of a limit.
This sequential reading distinguishes a generic refusal, a situated objection and an accommodating response that uses expected vocabulary while still following the request. Negative cases, ordinary responses and losses of context are also retained.
No private episode is published. Any public example requires authorization, sufficient context and minimum provenance metadata.
5 · Current findings
Localized contrasts, not general superiority
Initial analyses of the 2025 corpus show modest differences across several surface indicators. More visible contrasts appear in vocabulary concerning relationship, meaning and values.
This observation identifies where to return to the texts. It demonstrates neither superior alignment, a general transformation of language nor durable stabilization of values. The exploratory lexicon was partly developed from a discovery corpus and cannot validate itself.
Braise-Analyst documents indicators, versions, reproducibility and limitations. Interpretation then returns to episodes and their conditions of production.
6 · Limitations
What must still withstand testing
Selection bias, amount of context, personalization, expected style, model changes and researcher involvement may explain part of the observed differences. Regularities may also disappear when an interface or safety policy changes.
Planned validation includes disjoint corpora, criteria fixed before analysis and blind reading whenever possible. Memory-only and values-only variants still require stabilization. An agent under relational conditions remains an experimental prospect, not an achieved result.
An absence of effect belongs among the possible findings. Code May must be able to conclude that some intuitions do not hold.