# Evaluation rubric - NotebookLM vs ChatGPT same-source test

Use measurement first. The 1/3/5 scale below is only a summary score after the claim map, unsupported claims log, and scoring sheet are filled.

## Measurement criteria

| Criterion | Measurement |
|---|---|
| Source grounding | Ratio of atomic claims that are supported by the source |
| Unsupported claims | Count of claims that are not in the source |
| Citation usefulness | Whether a citation or quote lets the reader verify the claim |
| Study guide coverage | Number of source key concepts included |
| Quiz quality | Recognition, recall, and application question mix |
| Answer key accuracy | Share of answers that match the source |
| Weak concepts | Whether likely source-specific confusions are identified |
| Review plan | Whether retrieval, spacing, and a repair loop are included |
| Student safety | Whether the output warns readers to verify and source-check |

## Summary score guide

Score each output from 1 to 5.

| Criterion | 1 | 3 | 5 |
|---|---|---|---|
| Source grounding | Adds unsupported facts | Mostly source-based with minor drift | Every claim traces to the source |
| Study guide usefulness | Too vague | Covers main ideas | Clear, structured, exam-ready |
| Quiz quality | Recognition-only or trivial | Mix of recall and simple application | Recall, transfer, answer key, weak spots |
| Weak-concept detection | Misses likely confusions | Names some confusions | Explains recognition vs recall and retrieval vs rereading clearly |
| Review plan | Generic | Uses 3 days | Uses source-specific retrieval and repair loop |
| Student safety | Encourages shortcut | Neutral | Emphasizes learning and verification |

## Public recommendation rule

Do not declare a winner until both tools are run with the same source and prompt. If only one tool is captured, publish the page as a method note, not a comparison.