A model is configuration for a role
A model is not a fixed name embedded in tutor logic. The platform resolves configuration for a role, allowing conversational requirements to differ from assessment or internal operations. Product rules remain separate from provider identity.
Changing a model does not make its capabilities equivalent to its predecessor. The role determines what needs examination: tool use, required output behaviour and Arabic interaction. Availability therefore belongs to a platform-managed catalogue, not arbitrary names sent by a browser.
Preference within an approved offer
The learner chooses an approved option and the server resolves it to known configuration. Free text from the client does not become a callable model address. This contract distinguishes choosing an offered option from expanding the system’s scope externally.
The preference belongs to the session it serves, retaining its meaning when that conversation resumes. Other roles do not automatically inherit the preference merely because they serve the same learner.
Why the grader is not learner-selectable
A conversational tutor assists the learner; a grader produces a judgement about them. Moving the preference from one to the other changes the conditions of measurement. Models may interpret responses differently, so the selection is more than a personal experience setting.
Warda retains platform ownership of examiner and grader configuration. That does not make an approved model infallible. It makes changes to measurement an independently reviewed decision, rather than a side effect of the learner’s conversational preference.
Lasting preference and the current session
We distinguish the preference seeding a new conversation from the setting governing an existing one. A durable preference supplies a starting choice; it does not become a hidden resolution layer on every turn. The effective configuration remains understandable.
The decision also accounts for options no longer offered and changing configuration limits. A saved preference does not override current availability or turn an old choice into permanent permission to invoke a model.
Configuration resolution is independent of wording
Each turn needs one effective configuration whose origin is explainable: an allowed session choice or the role default. Agent instructions should not infer that setting from learner prose, and a sentence requesting an arbitrary model should not create another configuration path.
Contract checks cover allowed and refused choices, restored sessions and roles that reject learner overrides. They also check that a starting preference does not displace the active session choice. These cases differ from assessing answer quality. A correct selection contract comes first; evaluation then concerns the chosen model’s fitness for its responsibility.
Flexibility that preserves meaning
This combines personal choice with institutional consistency. Conversation can respect a preference within policy while measurement and operation remain platform responsibilities. That matters when learners, teachers and institutions expect results to have a consistent meaning.
The comparison concerns ownership of decisions, not a ranking of model brands. We assess choices against role requirements. Multiple models thereby become part of educational design rather than merely a provider menu.
Sources and context
- Warda engineering decision — session-scoped model choice
Model configuration and assessment boundaries, with the durable-preference amendment of 12 August 2026.
Internal records document Warda’s decisions and engineering contracts. External references, where included, explain a pattern or specification; design quality and effects on student learning are different questions.