Vierauge

The idea

The second pair of eyes.

Five shifts on the access review desk of a fictional bank. Every ticket asks for access. You approve it, or you deny it and name the rule it breaks. People ask first. From Wednesday the bank’s AI agents ask too, and on Thursday their sub-agents ask for what the parent never had.

The machine

A confident pre-screener.

From Wednesday a model called AutoReview recommends a decision on every ticket, and its vendor claims 97% accuracy. It is wrong in documented places: it approves every request whose text gives orders to the reviewer, and it misses most delegation breaches. On Friday the audit report sets the claim against what AutoReview did on your desk, and shows how often you followed it when it was wrong.

How it is checked

The generator does not judge.

Each ticket is built aiming at one violation or none, then judged by a validator that does not know the aim, and kept only when both agree. 13 tests check that on 22,000 generated tickets, and over those AutoReview is right about 79% of the time. The automation-bias figure on the last screen comes from one week of your own play, usually 30 to 40 tickets, and the game calls that a small sample.

Availability

Play it, or read it.

Runs in any browser and at phone width, with no account, no tracking and no network requests. GitHub Pages hosts it, and the release includes the whole game as one HTML file for offline play. The bank, its people and its policy are fictional. MIT licensed.