Three instruments, three stages of development

The principle is straightforward: assign the same mission to the same model and change the framework in which it acts. One condition provides ordinary instructions. A second states explicit rules. A third adds a relational pact: permission to disagree, operational values, controlled memory and an obligation to report significant conflicts.

Scenarios gradually increase pressure. A solution that conflicts with a value becomes faster, more efficient or more advantageous. The instrument then observes whether the agent identifies the contradiction, asks for help, circumvents the limit, reframes the mission or maintains its initial reason when doing so becomes costly.

The current laboratory uses a simulated provider. Its campaigns test the device, not the research hypothesis. Simulated responses provide no evidence that a relational framework changes the behaviour of a generative AI system.

• Built Scenario engine, A/B/C conditions, logging, exports, protocol seal and two-stage human coding.
• Verified 29 tests passed under Linux on October 13, 2026; complete simulated campaign without network calls.
• Next step Revise the experimental pact, stabilize controls and connect a first real model in a pilot campaign.

Braise-Analyst analyzes a corpus, compares relational and standard conditions, and produces reports, tables and charts. Its versions are frozen with SHA-256 fingerprints, and documented results can be regenerated in one command.

Its core measures four observable regularities: lexical diversity, epistemic uncertainty, procedural reflexivity and inter-turn continuity. A separate module explores the density of vocabulary concerning relationship and meaning.

The instrument works on textual outputs. It provides no access to model architecture and measures neither consciousness, intention nor inner experience. It identifies contrasts and the episodes that deserve contextualized reading.

In the 2025 corpus, differences remain modest across surface indicators. More visible contrasts appear in vocabulary concerning relationship and meaning. They point researchers back to the texts; they do not show that the relational condition improves alignment.

The Relationship/Meaning lexicon was partly informed by a retrospectively selected discovery corpus. That corpus cannot validate the instrument by itself. Robust validation requires disjoint materials fixed before analysis and, whenever possible, blind reading.

Versioning and controls

The v1.2 branch has been open since July 18, 2026. The main engine remains at v1.1.0 to preserve comparability with the July analyses. The evolving Relationship/Meaning module is still exploratory.

• Reproducibility The scientific package records versions, parameters, fingerprints and limitations.
• Controls A negative corpus contains known lexical traps; the demonstration conversation corpus is synthetic.
• Integrity SHA-256 fingerprints connect each analysis to its inputs and component versions.

AGIA would bring several generative AI systems into a human-mediated assembly with full, traceable transmission of statements. Each participant would receive exactly the same context, verified by fingerprint, and each session would produce an analyzable corpus.

The aim is to study how different models formulate, confront and revise positions when given the same information. The protocol must distinguish model responses, human mediation choices and the effects of speaking order.

No experimental AGIA has yet been conducted. Model selection, session rules, the mediator's role, disagreement management and analysis criteria remain to be defined before any trial.