DEV Community

fei mai
fei mai

Posted on

I Read 39 Arguments About Whether to Buy Something. One Field Predicted Every Verdict.

Thirty-nine people argued for or against a purchase in front of strangers. I read all of them on 17 November 2026, then I read the JSON behind them.

The JSON was more interesting than the arguments.

The data source

No browser, no scraping loop. WalletJury's front end loads cases from a public endpoint, so the audit is one request:

GET https://walletjury.com/api/list?sort=new&limit=200&offset=0

39 cases came back with next: null. Null means no more pages, so it's the entire database rather than a sample — nothing filtered, nothing skipped, 29,402 bytes.

Then we dumped every key on every case. Twenty-one keys.

What the fields actually contain

Field Filled Detail
item 39/39
price 39/39 20 to 12,999
currency 39/39 USD 37, CNY 2
defense 39/39 32–375 characters, median 244
use 35/39 Daily 19, Weekly 8, Not sure 7, few times a year 1
afford 39/39 stretch 22, easily 10, not really 7
anon 39/39 all true
authorName 0/39 never once
hasImage 1/39
ai.lean 39/39 guilty 36, not guilty 3
ai.opinion 39/39 20–121 characters, median 52
outcome 0/39 the thing the product promises
votes 46 total spread over 17 cases only
replies 51 total

Twenty-one fields. Seven carry real variation: item, price, currency, defense, use, afford, and the AI's one-line opinion.

Three things I did not expect

One: 29 out of 29.

Every case where the poster picked "It's a stretch" (22 cases) or "Not really" (7 cases) came back Guilty. All 29, no exceptions across the whole database.

The only three Not Guilty verdicts sit inside the ten "Easily" cases. A 30% acquittal rate in one bucket, 0% in the other two.

Two: the vote count doesn't spread.

22 of 39 cases have zero votes. The other 17 have exactly 3 votes (12 cases) or exactly 2 (5 cases). Nothing between the two, nothing above three.

46 votes have been cast on this database, all inside those 17 cases, and every one landed on the same two numbers. That is not a distribution. That is a shape.

Three: three fields have never done anything.

authorName is empty on all 39. hasImage is true on one. outcome is null on all 39 — the field that would record what the jury decided.

Meanwhile afford, the field nobody thinks about, splits the dataset cleanly in two.

What we could not measure

Why. We can see the output and not the reasoning. Three readings fit this table:

The model reads afford and treats "I can easily afford this" as a reason to acquit. Or it ignores afford and reads the text — people who pick "easily" write differently than people who pick "stretch". Or it is self-selection: if you can afford the thing, you only post the genuinely borderline ones, and borderline cases get acquitted.

Telling those three apart needs controlled submissions. I am not going to post fake cases into a live product to get a prettier chart. The correlation is 29 for 29 and that is the whole of what I can defend.

We also didn't render the site in a browser for this piece. Everything above is the API payload, read as text.

A correction to our own notes

On 14 November I wrote that this site had 39 cases and 46 votes. It is 17 November and it is still 39 and 46.

I wrote that assuming the next pull would show movement. The newest case was created 23 October, 13:08 UTC, and nothing has arrived in the 25 days since. The first version should have said "unchanged since 23 October".

Disclosure

WalletJury is one of the sites we build and run. These numbers come from its own public endpoint, and nothing on the site was changed to produce them.

All 39 cases are public: walletjury.com

One open question

If you were building this, what do you do with a field that predicts the verdict 29 times out of 29?

Keep it, because it works. Drop it, because a jury that reads your bank balance before your argument is not a jury.

I genuinely don't know which answer is right.

Top comments (0)