Export your reviews as a CSV, or take the public companion file from the book's data repository. Open a Google Colab notebook with Gemini enabled, upload the file, and paste the request below. This is a chapter 3 environment doing chapter 2 work: before you model anything, find out what you are holding.

Companion file and code: github.com/jeong-yeol/AOR_with_AI. Maple City v3 is 11,010 synthetic reviews across 200 restaurants, 2019 to 2022, and every number printed in the book is computed from it.

1. Check the row count against what you exported. A gap means the platform, the export, or the load dropped something, and you want to know which before anything else.

2. Look at reviews per business. If the median business carries a handful of reviews, no method in Part II will rescue that, and the honest move is to change the unit of analysis rather than the estimator.

3. Read the date range as a history of the platform, not of the market. Does your window cover the change you care about, with enough on both sides of it to see a before and an after?

4. It named columns it would not trust. Take one and decide for yourself: is it merely noisy, or is it measuring something other than what its name promises? Those two problems have different fixes.

Nothing on this page shows what any of these requests returned, and that is deliberate. The output that teaches you anything is the one your own file produces; a printed answer would turn a piece of work back into a piece of reading.

Same notebook, same file. This is the chapter 5 move that every later chapter depends on, run as an audit rather than a demonstration. You are looking for the cases where the score and the star rating disagree.

1. Read the 15 disagreements and label them yourself before you accept the score. If most of them turn on sarcasm, negation, or a mixed review with one loud complaint, your variable is tracking wording rather than evaluation.

2. Correlation with the star rating is a sanity check, not a validation. Very near one means you have rebuilt the star rating slowly; near zero means at least one of the two is not measuring what you assumed.

3. Find the thin business-months. Every panel, time-series, and spatial method in Part II inherits them, so decide now whether to aggregate up, restrict the sample, or narrow the question.

4. Say what the number measures in one sentence without using the word sentiment. If you cannot, you have a column, not yet a construct, and the chapters after this one will amplify that gap rather than close it.

This is the chapter 13 discipline in its smallest form. The request asks for labels and nothing else. The scoring is yours, because a measure you did not check is a measure you cannot report.

1. Label the same 100 yourself first, ideally with a second person, and settle your own disagreements before you look at the machine labels. How often two humans disagree is the ceiling on what any classifier can be scored against.

2. Compare row by row and count the two kinds of error separately. A single overall figure hides the direction of failure, and the direction is what determines whether your downstream estimate is biased up or down.

3. Reread the cases where you and the machine disagreed. If your category was ambiguous on those, fix the category definition and rerun, rather than blaming the classifier.

4. Write the sentence you would put in a manuscript, with the sample size and the agreement statistic in it. If that sentence needs a hedge, the hedge belongs in the manuscript too.