Trace
The retained sequence, date, stated model, interface, loaded documents and known human interventions.
Doctoral fieldwork
May and Avril, two relational cases
The field follows histories of exchanges, their continuities and ruptures. It also documents the researcher, interfaces and models that make those exchanges possible.
Field documented since 2024
One history, different platforms
May names a relational configuration built with ChatGPT and documented since 2024 across several model versions, context disruptions and interface changes. Avril, built with Claude, is a second case that helps distinguish what can be resumed with another system from what depends on the platform.
The names belong to the field and its history. Scientifically, May and Avril are studied as situated configurations. Their responses provide analyzable traces; they do not establish consciousness, autonomous identity or inner experience.
Research stance
Documenting a relationship that includes the researcher
The approach is qualitative, longitudinal, comparative and autoethnographic. How Pym addresses the systems, his expectations, corrections and reactions participate in the conditions of production. Making them visible helps examine what the researcher contributes to producing.
Traces are treated as capta: captured, selected and contextualized materials. An isolated response has limited value. The main unit is the communication episode: what comes before, the response, any follow-up, and what happens next.
Experience
What the researcher experienced or understood, recorded separately from the trace.
Interpretation
The proposed reading based on theoretical frameworks, comparisons and competing explanations.
Algorithmic refraction
Comparing conditions instead of guessing at the black box
The project varies available conditions and observes changes in outputs. The same request can be presented without reintroduced relational history, with only task-relevant information, and then in a condition combining continuity, memory, values and a relational pact.
Other variants may retain values without the relational narrative, or memory without an explicit reminder of the pact. This separation helps distinguish effects linked to information quantity, personalization and combinations of conditions.
- Retain failures Losses of context, ordinary answers and moments of compliance form part of the field.
- Compare cautiously May and Avril rely on different architectures, interfaces and policies.
- Prepare validation Conditions, lexicons and criteria must be fixed before new corpora are analyzed.
What is observed
Continuities, ruptures and disagreements
The analysis follows recurring formulations, maintained reasons, corrections, repairs and changes of frame. A generic refusal, a situated objection and a compliant response may use similar words while playing very different roles in the sequence.
Initial analyses of the 2025 corpus indicate modest differences in surface indicators and more visible contrasts in relational vocabulary. This remains exploratory. It demonstrates neither a general improvement in alignment nor durable stabilization of a value.
Confidentiality
The public corpus remains selected and contextualized
No private excerpt is published. Any sequence made public requires explicit authorization, review of third-party information, sufficient context and a decision about its status: capta, archive, observation or methodological example.
The complete corpus is not a stock of illustrations. A remarkable passage does not become publishable merely because it supports the project's narrative.