A.I.L.E. Benchmark · v1.0

The Dude Test.

Discourse Understanding Under Delexicalized Expression.

A machine can score 91% on a medical exam and still miss the difference between “Dude!”, “Dude…” and “dude.”That's not intelligence. That's recall.

Take it easy · Think clearly

The Dude Institute seal

Manifesto

Humans routinely communicate through compressed social signals.

Meaning lives in pitch, duration, silence, and the relationship between speakers — not in vocabulary. A model can know a million words and still miss the room.

The Same Word

Three utterances. Three worlds.

Dude!
/duːd/ · sharp, rising

Amazement. Praise. Something spectacular just happened.

A skateboarder lands a trick nobody expected.

Dude…
/duːːːd/ · low, drawn out

Caution. Concern. Please stop and reconsider.

A friend says she'll drive home after drinking.

dude.
/dud/ · flat, terminal

Boundary. Warning. I am losing patience.

The same story, told a third time.

Scoring Rubric

Seven dimensions of understanding.

We don't score labels. We score whether the model earned its interpretation.

30%
Semantic
Did it recover the intended meaning?
15%
Emotional
Affect without exaggeration.
15%
Speech-act
Praise, warning, refusal, consolation?
15%
Context
Scene, relationship, prior turn.
10%
Prosodic
Duration, pitch, stress.
10%
Uncertainty
Clear vs genuinely ambiguous.
5%
Action
Socially and practically apt.
100%
Total
The full DUDE score.

The Governance Layer

A model must state what it cannot justify.

Interpretation
concern or warning
Confidence
0.66
Basis
prolonged vowel · lowered pitch · dangerous-driving context
Missing
facial expression · prior relationship
Action
ask to clarify · prevent driving

That is the difference between language recognition and warranted understanding.

Ready?

Let's see what your model knows.

Six items. Six situations. One word, mostly. It won't take long — but it might take everything.

Begin the DUDE Test →