Evaluate before expanding
First complete the teaching lesson. Use the PocketCards codelab and this source revision for the practical work.
Your task
Create ten source passages and expected facts before running the model. Include sparse, contradictory, repeated and instruction-like notes.
Evidence to produce
Submit supported-answer and fact-coverage counts, failure observations and one justified prompt change. Record actual model and source revision. Keep a small learning log in your own branch. Include the command or action you used, the expected result, the observed result and the reason for any difference. Do not include configuration values, tokens or private source notes.
Before checking the answer
Predict the outcome before you run the app. If the result differs, locate the responsible layer: input validation, ViewModel state, provider adapter, storage or rendering. Fixing the wrong layer can hide a symptom without meeting the requirement.
Review guidance
A useful evaluation reports failed requests and denominators. Valid format and attractive prose cannot establish factual support.
Explain your decision
What would make you reject a prompt change even if average latency improved? Answer with a concrete code reference and one observation from your run. If you can only repeat the terminology, revisit the lesson and demonstrate the behavior before moving on.
Check your work after attempting the exercise
For every generated answer, identify its supporting source statement or mark it unsupported. Count distinct target facts for coverage. Keep failed requests in the results and report their outcome separately from answer-level ratios.
If your result differs: If all ten cases were chosen after inspecting outputs, rebuild the set before making a quality claim. An empty result has no answer-quality denominator; do not label it 100% accurate. Preserve difficult cases when comparing prompts.
Ready to continue: Your learning log contains a prediction, an observed result and a source reference. You can explain the result without copying this guidance. Carry the evidence into the final assessment.