DEV Community

fei mai
fei mai

Posted on

I Read My Own Voting Site's API and Found 46 Votes, No Verdicts

A voting site where nothing ever gets decided is a bug, not a feature. I found that out by reading my own product's JSON endpoint instead of looking at it.

The first case on WalletJury went up on 15 October. A $24 oversized wool-blend coat. Then nothing for sixteen hours, and at 08:16 the next morning ten cases arrived inside a single minute.

I did not spot that by watching the feed. I spotted it by dumping the table and sorting on createdAt.

Here is the setup. WalletJury is mine. Users post a purchase they are about to make, everyone else votes Guilty or Not Guilty. The rendered page is 98 KB of JavaScript around 169 characters of visible text, so scraping the HTML gets you a currency selector and nothing else. The real data lives at /api/list?sort=new&limit=200, which returns all 39 cases in one GET with next set to null.

So on 14 November I fetched that endpoint through a local proxy, saved the payload, and did the counting in Python. No sampling. The sample is the database.

What I counted Result
Cases in the database 39
First / last case 15 Oct 16:42 / 23 Oct 21:08
Days since a new case 21
Votes, guilty / not guilty 33 / 13
Cases with no votes at all 22 of 39
Most votes on any single case 3
Cases with a recorded outcome 0
Cases posted anonymously 39 of 39
Cases with an image 1
Model pre-verdict: guilty 36 of 39
Jury result across the 17 voted cases 16 guilty, 1 tie, 0 not guilty

Three findings, in the order they annoyed me.

One. outcome is null on every single row. Thirty-nine cases, seventeen of them carrying a completed 3-0 or 2-1 tally, and not one recorded verdict. Either a closing threshold exists and nobody reached it, or the closing step shipped as a TODO. I can see the column is empty. I cannot see the server, so I will not pretend to know which.

Two. The crowd has never acquitted. Sixteen guilty, one tie, zero not-guilty. Strong-looking signal until you check the model's pre-verdict: guilty on 36 of 39, and guilty on all 17 cases that later drew votes. A classifier that answers "guilty" nine times out of ten will appear to agree with humans at almost any sample size. The agreement is not evidence of judgement. It is evidence of a prior.

Three. None of the voters have accounts here. The replies field totals 51, and all 51 belong to the 17 voted cases, exactly three apiece. Every one of those cases carries a last reply object tagged with a source: facebook 7, reddit 6, instagram 3, x 1. Thirteen distinct handles across seventeen replies. Zero attributed to a WalletJury user.

And a detail I keep thinking about. In 12 of those 17 cases, the most recent reply argues not-guilty while the counter reads guilty. The visitor's last impression is the losing argument.

What I deliberately did not do: guess whether three is a hard cap. The API returns two integers, g and a, with no vote timestamps and no voter IDs. A cap of three and three coincidental visitors are indistinguishable from the outside. Writing that up as a finding would have meant inventing the part I could not see.

I also skipped the question of why posting stopped. Three cases a night at 21:0x from 17 to 23 October, then silence. I did not read the cron before publishing, so I am not going to narrate a cause.

A correction while I am at it. On 10 November I described this site as having "39 cases and 46 votes, and it is growing." It has not grown. Same 39, same 46, four days later. I looked at the total and skipped the timestamps, which is the same shortcut that made me wrong about a crawler last week.

Disclosure: WalletJury is one of my sites, which is the only reason I can hit the raw endpoint. Front end is at walletjury.com. I did not vote, post, or reply on anything during measurement — stuffing the ballot would have made the only number that matters meaningless.

The most recent case is a $99.99 Minecraft creeper mini fridge. Zero votes, zero comments, no verdict, twenty-one days old.

Top comments (0)