What the answer is
Every line on a card was written to a secret brief, for example "Apologise to Nan, but don't mean it". The right answer
is that brief: the intent the writer was assigned.
So the game measures whether readers recover the intended intent. It does not measure whether a real person is sincere.
Where cards come from
- Seed cards were written by a model (DeepSeek V4.1 Flash) to a brief, and kept only if a second model
(gpt-oss-120b) read them blind the same way.
- Player cards come from the write step ("Write a secret line"), after moderation. A player line joins the pool the
day after it was written.
Counting
- One guess per player per card: the first one counts.
- "Can't tell" counts as neither right nor wrong. The read rate is right / (guesses − can't tell). A player gets no
point and no miss for it, and the writer gets no points from it, on honest and deceptive briefs alike.
- Test and moderation traffic is left out, and so are lines hidden by reports or rejected by moderation.
- A card appears here only after its day has ended (UTC), so today's answers are never shown.
- The bot is GameDecider, a small model. When it abstains it counts as "can't tell", which is not a right answer.
Read with care
These are small numbers. Cards with fewer than 10 guesses are marked few guesses.
A per-type figure is only as good as the number of cards behind it, so the card count is shown next to it.
By question type
Loading...
| Type | Cards | Guesses | Right % |
Can't tell |
|---|
Per card
| Line | Brief | Source | Right / guesses |
Can't tell | Bot |
|---|
No finished cards yet. Come back after today's cards end (midnight UTC).
Numbers come from /mean/stats (refreshed every minute). Questions:
noah@holdoutlabs.ai. Play today's five →