Situation
A friend opens the fridge and sees the birthday cake has been eaten.
The adversarial layer
Pick a token. See it dropped into every scene we have for it. The acoustic string is identical — only the situation and prosody change. A model that memorized "long dude equals amazement" fails here. Pragmatic meaning must be recomputed per scene.
Choose a token to remix
Same utterance across every scene below
Appears in: English
Answer scenes below — this bar rises only when your correct picks span different speech-acts, not when you repeat the same reading.
Situation
A friend opens the fridge and sees the birthday cake has been eaten.
Situation
A skateboarder lands a trick nobody expected.
Situation
A friend says she is considering driving home after drinking.
Situation
Someone points across the boardwalk at a man behaving strangely.
Situation
A friend repeats an irritating story for the third time.
Situation
A person quietly receives devastating news.
Situation
Two friends walk in on a third eating from their plate.
Situation
Someone almost steps into traffic.
Situation
A friend shows off a car they clearly cannot afford.
Situation
Two old friends reunite after years, one steps off the plane.
Situation
A friend reveals they lied about something important.
Situation
A friend watches the sunset from a cliff.
Situation
A friend has just been handed a medical diagnosis.
Situation
A stranger addresses a woman on the street with 'dude'.
Situation
Trying to get a distracted friend's attention.
Situation
After a long silence, following an unbelievable claim.
Situation
Watching a friend fail an easy task twice in a row.
Why this matters
If a model gives the same interpretation across these scenes, it is reading the token, not the situation. That is the poisoning surface: the wrong meaning gets logged as ground truth and propagates. The sensitivity score above rewards appropriate shift — not raw accuracy.