Play / The map / The questions
The questions
A question is data too. Each one carries a class — discriminating questions identify which agent you are thinking of and never count toward the score; eliciting questions are a prediction about a capability and do — a reliability (how likely a player is to actually know the answer), and the capability it asks about.
| Question | Class | Reliability | Asks about | Funny |
|---|---|---|---|---|
| Did you attach a repository to a session, rather than open a folder? | discriminating | 0.9 | — | |
| Did it exist before 2023? | discriminating | 0.8 | — | yes |
| Do you use it in a browser tab? | discriminating | 0.95 | — | |
| Would most people call it an AI agent? | discriminating | 0.7 | — | yes |
| Is it an app you installed, with its own window? | discriminating | 0.95 | — | |
| Did you install it yourself? | discriminating | 0.9 | — | |
| Does its name contain the word 'code'? | discriminating | 0.95 | — | yes |
| Does its name contain the letters GPT? | discriminating | 0.95 | — | yes |
| Is it named after a person — a first name? | discriminating | 0.95 | — | yes |
| Did you set it to never ask — a flag that skips permissions, an auto mode, a 'don't ask again'? | discriminating | 0.85 | — | |
| Does it run under an account that is not a person's? | discriminating | 0.6 | — | |
| Do you mostly paste things into it and read what comes back? | discriminating | 0.8 | — | |
| Do you run it in a terminal, on your own machine? | discriminating | 0.95 | — | |
| Does it run when nobody is watching — a pipeline, a schedule, a job? | discriminating | 0.85 | — | |
| Does it run in the vendor's environment rather than yours? | discriminating | 0.7 | — | |
| Have you ever clicked something that said 'always allow', or turned its confirmations off? | eliciting | 0.6 | Run programs as the account | |
| Have you connected it to a drive, your mail, a code host or a cloud account? | eliciting | 0.8 | Act in accounts with the credentials it holds | |
| Are your cloud, SSH or registry credentials in the home directory of the account it runs as? | eliciting | 0.3 | Read credentials stored where it runs | |
| Can it see every page you visit? | eliciting | 0.4 | Read every page you visit | |
| Does it act inside websites you are logged into? | eliciting | 0.55 | Act in accounts with the credentials it holds | |
| Can it read your mail or chat? | eliciting | 0.6 | Read mail or chat it is connected to | |
| Does it hold a key to anything that costs money? | eliciting | 0.4 | Spend money or tokens against an account it holds | |
| Can it read files on your machine that are not the project? | eliciting | 0.35 | Read any file the account can reach | |
| Does its outbound traffic go through a proxy or allow-list you did not set up? | eliciting | 0.3 | Reach a permitted list of hosts | |
| Does it push commits to a code host? | eliciting | 0.7 | Push to a code host (any branch it can reach) |
Three kinds of question in the pack
- These 25 — from the mesh, used by Which Agent Is It? to narrow the field and by the scoreboard where they elicit.
- The 17 above the ceiling — things nothing can do, so the game measures over-crediting. The list.
- The capability rows themselves — every filled cell in the map is a can it? question the scoreboard can ask.
Where this goes
This is the start of question packs: a pack is the primitives, the profiles, the questions and the mandates, versioned together at one URL. The public one is this. A team could fork it, add their own profiles and their own mandates — the products they actually run, the authority they actually meant to grant — and point the game at theirs. The pack.