# American Flag VADER Known-Effect Evaluation

This citable artifact evaluates whether Conjoint Survey's VADER text-analysis workflow recovers a deliberately engineered sentiment difference in a two-arm synthetic survey experiment.

## Plain-language design

- Control: no image, followed by the prompt "What does being American mean to you?"
- Treatment: American flag image, followed by the same prompt
- Sample: 1,000 synthetic respondents, evenly divided
- Known pattern: treatment responses were intentionally generated with more positive language; control responses were generated with neutral and mixed language
- Outcome: VADER compound sentiment from -1 to +1

## Result

**Passed.** Treatment mean 0.543, control mean 0.181, difference 0.362, 95% CI 0.325 to 0.400, randomization-inference p < 0.001.

## Critical limitation

This artifact does **not** show that exposure to an American flag makes real respondents more positive about America. The response generator intentionally creates the difference. The evaluation asks only whether the scoring and comparison pipeline detects a known difference when one exists.

## Files

- `evaluation-report.md`: readable report
- `evaluation-report.json`: machine-readable results
- `synthetic-responses.csv`: response-level synthetic data
- `study-design.json`: instrument and estimand
- `american-flag.svg`: treatment stimulus
- `codebook.md`: variable definitions
- `replication.mjs`: independent score-and-compare script
- `CITATION.cff`: citation metadata
- `metadata.json`: repository and archive metadata

## Reproduce

From the Conjoint Survey repository root:

```bash
npm run validate:flag-vader
node docs/research-artifact/american-flag-vader-evaluation-v1/replication.mjs
```

No human records or personal data are included.