The Polyglot Challenge

You claim you speak ten languages. A conversation lets you bluff. A dictation doesn't. Prove it statistically — I'll go first.

There is a man on a stage. He speaks ten languages — the bio says so, the video titles say so, the applause agrees. Watch enough of these performances and you notice something underneath the applause: the definition of "fluent" quietly changes with each language he adds. For the first two it means conversation with natives at full speed. By the seventh it means ordering food and a rehearsed paragraph about how he learned it. Nobody checks, because there is nothing to check. The claim has no scoreboard.

This post is a scoreboard.

Why the claim survives: conversation is bluffable

A live conversation is the friendliest possible test of a language. You have context. You have the other person's face, hands, and tone. You can steer toward islands you've rehearsed and nod through the water between them. You can deploy the forty phrases you own with excellent pronunciation and let the listener's politeness fill the gaps. None of this is dishonest exactly — it's what skilled communicators do. But it means a conversation measures charm and preparation at least as much as it measures the language.

Verification is not generation. Recognizing a sentence when you can see the transcript, or half-knowing it from context, is a different act from resolving it out of raw sound with nothing to lean on. The ear that can only verify feels fluent right up until the moment it has to generate.

A dictation stream of random native-speed sentences removes every prop at once. No context. No face. No steering — you don't pick the sentences. Either the ear resolves the signal into the exact words or it doesn't, and the diff is character-level. There is nowhere to put the body language.

The chess argument

Chess solved this problem a century ago. Nobody on a chess stage claims to be "basically a grandmaster" — because ratings exist, and a rating is earned in open play where you don't choose your opponents. The pairing algorithm doesn't care what openings you prepared. Play enough rated games and the number converges on the truth about you, whatever you'd prefer it to be.

Dictation over a large corpus works the same way. The corpus deals you sentences the way pairings deal you opponents — randomly, from the whole population, at native speed. And every property of a chess rating transfers:

A provisional rating proves nothing. Twenty games, or a few hundred sentences — the number hasn't converged. One hot month is one lucky game. The claim is the sustained rating over thousands of encounters.

You cannot farm weak opponents. No phrasebook sentences, no travel-domain comfort zone, no rehearsed islands. The corpus contains bureaucracy, weather, arguments, literature, and a woman saying something quickly with a dog barking behind her.

Ratings decay honestly. Stop playing and the number drifts down — visibly, in the record, instead of in the fog of "well, I used to be fluent."

The polyglot who keeps redefining fluent is the player claiming 2200 strength who only posts wins against friends.

The metric

Success-New: your first-attempt success rate on sentences you have never heard before, drawn at random from a corpus of native recordings, graded by character-level diff. No warm-up, no familiarity, no second chances counted. It is the cold-start number — the closest thing to walking up to a stranger who has no reason to slow down for you.

The bar: a sustained Success-New of 70% or above in the language you claim. In practice, 70%+ cold-start on native audio corresponds to roughly C1 — the data behind that calibration is published here. Below that, what you have is real, and worth having, and not fluency.

Exhibit A: what an established rating looks like

I'm not asking anyone to take a test I haven't taken. Here is my Romanian, month by month, from the first session in March 2021 — 125,746 recorded attempts over five years. Every number below came from sentences the algorithm chose, not me.

Read it bottom-up. It starts at 39% — I was failing three new sentences in five. It stays under 50% for most of a year. That's the part the stage never shows: nine months of being demonstrably not good, in writing. Then the volume era — five to nine thousand attempts a month through 2022 — and Success-New crosses 70% and stays in the 70s for years. When I barely play for months, it drifts down toward the high 60s, visibly, like a rating should. It decays honestly and comes back with play.

That is what a claim looks like when it's real: not a number, a history.

Exhibit B: my own dismal table

Now here is my Portuguese, so you know the metric isn't rigged in my favor:

Success-New between 9% and 21%. Dismal. Also exactly correct: I'm a few hundred sentences into a language whose pronunciation seems designed as a practical joke on people who read Spanish. By the felt experience I was making progress — words were starting to surface, sentences occasionally snapped into focus. The table doesn't care how it felt. That's the point of the table.

The pair of exhibits is the argument. The same metric reads low where it should be low and high where it should be high, for the same person. It measures the skill, not the ego. My table can't lie for me either.

The challenge

So: you claim ten languages. Or five, or three. The claim is admirable if it's true, and checkable either way — that's new. No judges, no scheduling, no fee. The test is self-administered:

1. Open the app — it runs in the browser, no account needed to start. Pick a language you claim.
2. Play real sessions — twenty minutes of random native-speed sentences at a time, transcribed cold.
3. Look at your Success-New. Not your best sentence. Not your overall rate on repeats. The cold-start column, over enough volume to mean something.

Sustained 70%+: I believe you, and so does everyone you show the table to. Below that: also fine — now you know your rating, which is more than the applause ever told you. The corpus will happily supply the missing games. Every self-proclaimed polyglot with a TED talk is invited. So is everyone who ever sat in the audience wondering.

The tape doesn't care about your body language.