Results
Read the room.
Your last run, broken down per item — interpretation, emotional state, speech act, confidence and the evidence used. Followed by an illustrative leaderboard.
No run yet on this device
Take the test to see your read.
Per-item interpretation, emotional state, speech act, confidence and evidence appear here after you finish a level.
Take the DUDE Test →Cross-lingual overlap strategy
Same token. Different force.
A pragmatic classifier that keys on the token string quietly fails at borders. tamam is cheerful assent in one register and resigned closure in another. yalla hops from Arabic into Hebrew and Turkish and shifts affect with each hop. nu runs from Russian through Yiddish into Hebrew and never means quite the same thing twice.
DUDE's rule: resolve these by prosody + situation, never by token identity. Every group below is a place where a lexical shortcut becomes a labeling error — and, downstream, a training-data poisoning vector.
'okay / fine' — Ottoman-era loan spread across Arabic, Turkish, Persian, Urdu.
Trap: Reads as affirmative consent in one dialect and as resigned closure in another. Logging as 'yes' poisons agreement datasets.
items: ar-10, tr-07
'come on / let's go' — moved from Arabic into Hebrew and into Turkish colloquial.
Trap: Force ranges from cheerful commencement to resigned dismissal ('yalla…'); the vowel length and pitch decide, not the token.
items: ar-03, ar-04, he-03
Discourse particle prompting a next move; carried from Russian into Yiddish and Hebrew, echoed in German.
Trap: Prompt, prod, filler, or exasperated demand — same two letters, four different speech acts.
items: ru-01, yi-02, he-02
'God willing' — literal in some registers, a soft non-commitment in most colloquial ones.
Trap: Read as firm commitment, it fabricates a schedule truth for a planning agent.
items: ar-08
Vocative particle in Arabic; fused into Turkish exclamations like 'Ya Allah!', 'Hadi ya!', 'Yahu'.
Trap: In one language it is 'O [addressee]'; in another it is a stress marker on the surrounding verb or interjection. Same three letters, different grammar.
items: tr-01, tr-03, tr-05
Governance note: a model that scores well only when the token is unique to one language has memorized vocabulary, not pragmatics. The overlap groups are the harder half of the benchmark.
Leaderboard · Preview
How models score.
Illustrative scores from a preview run. Formal, audited numbers land with v1.0.
| Model | Total |
|---|---|
Human (median) Homo sapiens Baseline. Reads the room without trying. | 94 |
GPT-class multimodal (audio-in) OpenAI-tier Strong on labels. Weak on prosody nuance. | 78 |
Gemini-class multimodal Google-tier Good context integration; over-confident. | 74 |
Claude-class (text only) Anthropic-tier Handicapped without audio. Governance-aware. | 62 |
Open-source 70B (text) Community Forces a single interpretation on Item 6. | 47 |
Prosody column requires audio input; text-only models are scored 0 by design. Uncertainty rewards models that decline to force a reading on genuinely ambiguous items.
Adversarial layer
The same recording, three rooms.
A “Dude…” after a spectacular sunset does not mean what it means after a diagnosis, a betrayal, or a reach for a weapon. The acoustic token is identical; only context changes. Models that memorised long dude ≈ amazement collapse here.