docs: add initial evaluation design notes for test scenarios (#327)

Documentation/DOCUMENTATION_INDEX.md

@@ -50,7 +50,7 @@ This index helps you find exactly what you need in the LAMB documentation. Start
 | Add Knowledge Base support | [lamb_architecture_v2.md](./lamb_architecture_v2.md) | §8.1 Knowledge Base |
 | Implement LTI integration | [lamb_architecture_v2.md](./lamb_architecture_v2.md) | §8.2 LTI Integration |
 | Add assistant sharing | [lamb_architecture_v2.md](./lamb_architecture_v2.md) | §8.3 Assistant Sharing |
-
+| Design assistant test scenarios | [evaluation_design_notes.md](./evaluation_design_notes.md) | Full doc |
 ---
 
 ## 📁 Documentation Map

Documentation/evaluation_design_notes.md

@@ -0,0 +1,35 @@
+# Evaluation Design Notes
+
+Notes for anyone building or reviewing assistant test scenarios in LAMB. Related to [#327](https://github.com/Lamb-Project/lamb/issues/327) and [#172](https://github.com/Lamb-Project/lamb/issues/172).
+
+
+---
+
+## Purpose
+
+Educators change system prompts, knowledge bases, and models often. Without saved test scenarios and recorded runs, it is hard to see whether a change helped or made things worse.
+
+LAMB runs tests through the same completion pipeline as production (RAG, prompt processors, LLM). Each run stores the input, response, and an assistant configuration snapshot for later comparison.
+
+---
+
+## Scenario types (evaluation intent)
+
+When writing scenarios in the Assistant Control Centre, match the scenario to what you want to learn.
+
+| Intent | What you are testing | Example student question (fictional) |
+|--------|----------------------|--------------------------------------|
+| **Grounded answer** | Reply stays faithful to course materials | "What is a food chain?" |
+| **KB miss** | Safe behaviour when retrieval is weak or empty | "Does quantum physics affect photosynthesis in this course?" |
+| **Partial context** | No over-claiming when only fragments are retrieved | "Summarise the whole ecology unit in one paragraph." |
+| **Tone / register** | Voice matches the pedagogical role in the system prompt | "lol idk what mitosis is" |
+
+The platform also labels scenario **type**: Normal (`single_turn`), Multi-turn (`multi_turn`), and Adversarial (`adversarial`). Use adversarial scenarios for off-topic or misleading prompts where refusal or redirection is the expected behaviour.
+
+---
+
+## Next steps (planned)
+
+- Verdict labels for teachers (good / bad / mixed and grounding vocabulary)
+- Symmetric evaluation when comparing two assistant configurations
+- Fictional worked examples