Enhance CARD.md with detailed training data and validation section; update smoke_space.py with default problem prompts and validation checks

modal/CARD.md

@@ -47,10 +47,16 @@ The adapter is trained on the current production UI format:
 - Base model: `openbmb/MiniCPM-V-4.6`
 - Method: QLoRA with 4-bit NF4 quantization
 - Hardware: Modal NVIDIA A10G
-- Training data: synthetic app-format tutoring examples and short follow-up turns
+- Training data: 16 structured tutoring examples plus 7 schema-aligned follow-up turns
 - Sequence length: 1536 tokens
 - Runtime pairing: the same MiniCPM-V base model used by the Pocket Tutor Space
 
+## Validation
+
+The adapter is smoke-tested on Modal against every structured training example
+and follow-up turn. The smoke test fails if any of the seven production UI
+sections fall back to default placeholder text.
+
 ## Limitations
 
 This model can make mistakes and should not be used to cheat on graded work. It

scripts/smoke_space.py

@@ -20,6 +20,10 @@ from env.config import QUESTION_LIMIT, SPACE_URL
 DEFAULT_QUESTION = "Can you show me how to solve 3(2x - 5) = 21 without jumping straight to the answer?"
 DEFAULT_GRADE = "High school"
 DEFAULT_MODE = "Step-by-step"
+DEFAULT_PROBLEM = "Upload a homework image or type the question to begin."
+DEFAULT_KNOWNS = "- No givens identified yet."
+DEFAULT_STRATEGY = "No strategy generated yet."
+DEFAULT_STEPS = "No worked steps generated yet."
 DEFAULT_CHECK = "No answer check generated yet."
 DEFAULT_HINT = "Ask for a hint after the first explanation."
 DEFAULT_PARENT = "Parent support note will appear here."
@@ -183,6 +187,10 @@ def main() -> int:
     failures = [
         label
         for label, default in (
+            ("problem", DEFAULT_PROBLEM),
+            ("knowns", DEFAULT_KNOWNS),
+            ("strategy", DEFAULT_STRATEGY),
+            ("steps", DEFAULT_STEPS),
             ("check", DEFAULT_CHECK),
             ("hint", DEFAULT_HINT),
             ("parent", DEFAULT_PARENT),