Protectum

Polls in Serbia are rare, late and often doubted. A prediction market answers the same question differently: people put something at stake, so the price of a “yes” is the crowd’s real estimate. The app says it plainly on every board: the number is the price of a YES card, the sum of what people staked, not our estimate.

Two rules shaped everything else. Resolution is fixed when a market opens: the rule and the official source (for elections, the electoral commission) are written in advance and never change, and a correction only happens after a public objection within 24 hours. And there is no real money: everyone gets the same monthly demo salary, so the ranking measures judgement, not wealth.

I asked a simple question: do the AI markets even make sense? Six agents checked all 287 open AI markets against current news. 154 had a problem and 66 should not exist. Examples: “Will gold pass $2,700?” when gold was at $4,348. “Will inflation be above 4.5%?” after the statistics office had published 2.2%.

A second pass over all 774 open markets found 319 problems. 117 were voided and 1,890.79 demo € refunded. The root cause was the same every time: the writer does not know the current state of the world, and a single call cannot check itself. The prompt already said “no market if it has already happened”. It did not help.

Every draft now goes through two gates before a human sees it:

1. Mechanical checks, no AI, 11 unit tests. The deadline must be after the date in the rule, a threshold is compared against the current value, some words are banned, and there are no markets on court cases against private persons.
2. One small model call per draft with search. Is it already known? Does the deadline hold? Does the price match public polls? The answer is pass, revise or reject, with evidence and a date.

First round: 20 drafts, 3 passed, 6 to revise, 8 rejected (4 of them caught by the mechanical checks alone).

• Writer → controller → human → judge. The writer reads a sheet of news (RSS plus Bing News) and drafts markets; the controller checks each draft; a human ticks what gets published; the judge settles markets once the answer is public, and hands the uncertain ones back to a human.
• Edge functions have 150 seconds. Long model calls run in the background and a scheduled collector picks up the results, so nothing waits synchronously on a model.
• Everything scheduled with pg_cron, everything logged per run with token usage and price.

The first version logged nothing, so I added token usage and price to every run and put the daily cost on the admin home page. The real Google bill for the first week was $19.47. Then one day hit 97 calls and $12.61, while 736 AI markets had produced two real trades. I switched all six scheduled AI jobs off. Settling is now a manual button in the admin until the numbers justify turning them back on.

The iOS app went back to the home screen with no crash report. With temporary trace logs I found the cause: the feed rendered all 769 markets with images at once, JavaScript froze for about 4 seconds and iOS killed the app for memory. The fix was pages of 24.

The same code also runs on the web, as a board for the 2026 elections that anyone can cite: percentages, volume and how each market is resolved.

The video at the top was built only from the code. I rendered the native screens through react-native-web in a separate copy of the project, drove the real flow with Playwright (one real 1 € demo purchase), captured every scroll and slider position as its own frame, and assembled it in Remotion. Only the iOS status bar, the glass tab bar and the home indicator are drawn, because on a phone the system draws those too.