Your deliverable

Create a requirements evidence file containing an interview ID, an exact quote, a hypothesis and the next observation needed for each entry. The three interviewees are fictional. This exercise is not market research or proof of demand.

Download interviews.json, evidence.json, validate.py and the README. The sample interviews are in Chinese; the validator compares Unicode text without translating it.

1. Preserve the input

A manually copies booking requests; B is satisfied with an existing calendar; C is concerned about storing contact details. Preserve those differences. Do not summarize them as universal demand for automated booking. For real interviews, use authorized data, remove unnecessary personal details and retain a separate original.

2. Request evidence, not a product verdict

Use this prompt with the synthetic input in your chosen model:

Return a JSON array. Each entry has exactly these fields:
id, interview_id, quote, hypothesis, next_check.
quote must be an exact contiguous substring of the matching interview text.
Keep original-language quotes unchanged. State hypotheses as unproven.
Preserve contrary opinions. Do not invent frequency, revenue or willingness to pay.
One problem per entry. next_check must describe a specific question or observation.
Return JSON only. Omit claims that lack evidence.

The statement about copying requests supports investigating duplicate entry. It does not establish willingness to pay for a scheduling product. Observing one real request-processing session is a useful next step.

3. Check the citations locally

Save the downloads together. With Python 3 installed, run:

python3 validate.py interviews.json evidence.json

The completed example has three entries and should pass. Change a quote to a sentence absent from its source and run again: it should fail with a nonzero exit code. Restore the quote or remove the unsupported claim; never edit the source to fit the output.

The script checks structure, IDs and exact quote matching. It requires no API. Your model’s own usage charges are separate.

4. Review what matching cannot establish

A genuine quote can still be misleading in context. Review whether each hypothesis overreaches, whether contrary evidence remains, and whether the next check reduces uncertainty. A passing validator does not establish that a requirement is real or worth building.

5. Make a limited decision

For this synthetic example, investigate request collection before replacing everyone’s calendar. Keep B’s contrary account and separately investigate C’s storage concern. Completion means every hypothesis is traceable, the corrupted-quote test fails, contrary evidence remains and the next action is an observation rather than automatic implementation.

An unknown ID often means the input and output came from different batches. For malformed JSON, repair formatting without adding claims. For unauthorized or sensitive source material, return to the approved data-handling workflow.