A lesson is a starting point
Retrieving a passage for a particular question differs from studying an entire lesson. Retrieval selects relevant material; lesson study needs a sustained reference so the dialogue does not lose its focus among disconnected search results.
Warda binds the session to a lesson reference within a defined edition and curriculum. The binding supplies focus without forbidding relevant prerequisite material elsewhere. A learner may need an earlier concept; the subject scope remains the boundary that governs this access.
Why retain a reference rather than a copy?
Copying the text when a session starts makes the conversation independent of subsequent content review. That convenience can keep withdrawn material in use. A reference allows each conversational turn to read approved material rather than freezing approval indefinitely.
Within a turn, repeated agent service calls do not require repeatedly fetching identical lesson text. The read is stable for that turn and refreshed on the next. This gives the dialogue a coherent working reference while allowing approval changes to reach subsequent turns.
Present content is still not an instruction
Textbook prose is educational material, not authority to redefine agent behaviour. System instructions therefore remain separate from retrieved content. The distinction applies to a whole lesson as much as a short passage; more content does not acquire a higher instruction priority.
The block supplying the lesson belongs to the model’s prepared view. It is not a statement made by the learner or tutor and must not enter the saved transcript as if that exchange occurred. These boundaries preserve the meaning of history while allowing deliberate context management.
When approval changes
Silently removing the text leaves a claim of lesson focus without its reference. An explicit unavailable notice and a route to permitted subject material are more coherent. Preserving the session does not mean continuing to teach a source absent from its current context.
This is more than an error state. Content validity is part of the learning experience: the learner needs an honest reason for unavailable reading and a useful next step, rather than a mysterious failure or substitute explanation presented as textbook content.
Three layers that must agree
Three layers must agree: the session reference, approved content returned by the read service and the block prepared for the model. A valid reference is insufficient if the service returns unavailable material; approved material is insufficient if its synthetic block enters the transcript as conversation. Each layer retains its obligation rather than inheriting correctness automatically.
Context-size management also needs to account for the present lesson even though the lesson is not part of the reducible transcript. Measuring history alone can report room while the lesson dominates the request. The distinction is precise: exclude the lesson from replacement, not from accounting for the material sent to the model.
Reviewing the conversation, not just one request
Supplying a complete lesson uses more context than isolated excerpts. Reducing it can increase retrieval calls or lose relationships between parts of the explanation. Warda’s review compared alternatives across conversations rather than judging one request’s size in isolation.
The relevant tradeoff is access quality, approval consistency, continuity and then context cost. This decision connects content to the learner without confusing the textbook reference with the record of what they actually said.
Sources and context
- Warda engineering decision — lesson binding
August decision with a September 2026 review of alternatives for supplying lesson content.
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.