Verify this study yourself.

The numbers above come from published code, the same code this page runs. Download it from the Ballot Study V5 packet (the analysis engine and the reproduce script), upload it to an AI, and paste this prompt to have it look for any rule that inflates support.

You are reviewing the analysis code for a public survey for signs that it is rigged to inflate support. I have uploaded the analysis engine and the reproduce script (TypeScript): index.ts, parse.ts, filters.ts, stats.ts, waves.ts, types.ts, and rthg-reproduce.ts.

Read the code and answer plainly:
1. How is "support" defined? Is counting a 5, 6, or 7 on a 1-to-7 scale as support a fair, standard top-box definition, or is it gerrymandered?
2. Are the exclusion rules in filters.ts ordinary data-quality checks (attention, comprehension, speeders), or are any of them designed to drop people who oppose the measure?
3. Is the Wilson 95% interval in stats.ts implemented correctly?
4. Is anything hard-coded, faked, or short-circuited so the output does not actually depend on the input data?
5. Does any rule reference the respondent's vote or party when deciding whether to keep a row?

Give a candid verdict: is this analysis code honest, and would it produce a misleadingly high number?

Arithmetic that reproduces can still come from a leading instrument. Download the survey instrument and the pre-registration from the Ballot Study V5 packet (or read the methodology rationale), upload them, and paste this prompt to get a candid critique of question wording, order effects, and how "support" is measured.

You are a skeptical survey methodologist reviewing a study's instrument and pre-registration for bias, not its arithmetic. I have uploaded the survey instrument (rthg-study-definition.ts), the plain-English pre-registration, and the methodology rationale.

Critique the design directly:
1. Are any questions leading, loaded, or framed to push respondents toward supporting the measure?
2. Could the order of sections (rating the principles before voting on the law) bias the final vote? How much, and in which direction?
3. Are there demand characteristics or sponsor cues that signal a "right" answer?
4. Is the way "support" is measured (a 1-to-7 ballot, top-3 counted as support) fair?
5. Are the exclusion rules reasonable, or could they inflate support?
6. Is the sample (Prolific, nonprobability) being over-claimed as nationally representative?

Be candid. List the strongest legitimate criticisms a hostile expert would make, then say whether, on balance, the ~90% support figure looks like an artifact of the design or a real finding.

The strongest possible check is to run the study again yourself. Field the identical instrument to your own participants on your own Prolific account (so it costs us nothing), and compare your result to ours.

You can also take the study yourself and compare what you see, page for page, to the published instrument in the Ballot Study V5 packet.

The data contains no Prolific IDs. Each row has a salted pseudonym, so only you, knowing your own Prolific ID, can find your row, and no one's ID is ever exposed:

pid = "p_" + first 16 hex chars of sha256("rthg-public-pseudonym-v1" + YOUR_PROLIFIC_ID)

Compute yours and find your row (Python, no install beyond Python itself):

import hashlib, json
pid_input = "PUT_YOUR_PROLIFIC_ID_HERE"
mine = "p_" + hashlib.sha256(("rthg-public-pseudonym-v1" + pid_input).encode()).hexdigest()[:16]
print("your pseudonym:", mine)
for line in open("rthg-s5-dataset.jsonl", encoding="utf-8"):
    row = json.loads(line)
    if row.get("pid") == mine:
        print("found your row. demographics:",
              row["payload"]["pages"].get("section-g-demographics"))
        break
else:
    print("no matching row in this dataset")