Intellect: counting the words you did not repeat
By Noam Shemla · Data & Method
TL;DR Intellect scores the depth and lexical richness of what you said. Half of it is a model judging your argument. The other half is a type-token ratio (unique words divided by total words) computed from your transcript. Nobody votes on this axis, and the ratio has a known flaw you should understand before reading your own number.
Intellect is the newest of the five axes and the only one that did not exist in some form before. It scores two things at once: whether the ideas had depth, and whether the language carrying them was precise or generic.
Two signals, one axis
A language model reading the transcript, plus a type-token ratio computed deterministically from the same text.
Measured
- Type-token ratio (unique words ÷ total words)
Inferred, not measured
- The model's judgement of argument depth and originality
The ratio is mapped onto the 0–100 scale across a deliberately generous band:
No human in the loop
Intellect is the only axis with no human input whatsoever. That is a deliberate omission rather than an oversight.
Adding it would have meant a fifth star on the voting widget members use live, mid-battle, under time pressure. That meant a change to a real-time interface, a migration on two tables already full of votes, and an edit to every screen that reads the existing four-category shape. The judgement was that shipping the axis fed by AI and vocabulary alone was better than blocking it on a live-voting redesign.
One consequence: a member who only ever does Live Battles and Rounds has an Intellect score with no human corroboration in it at all.
The flaw in the ratio
A type-token ratio has a well-known property that makes raw comparison between texts unsound: it falls as a text gets longer. Every speaker eventually reuses the, and, is. A sixty-second clip will almost always produce a higher ratio than a five-minute one by the same person with the same vocabulary.
This is not a subtle statistical caveat. It is the central reason corpus linguists moved to length-corrected measures instead (McCarthy and Jarvis, 2010).1 We use the uncorrected ratio, which means the number is comparable across your own recordings only when they are of similar length.
The device it punishes Deliberate repetition is one of the oldest tools in rhetoric. Anaphora (beginning successive clauses with the same words) is the engine of the most-quoted speeches in English. A type-token ratio reads it as a small vocabulary. If you build a passage on a repeated phrase, expect this number to fall, and be right anyway.
What lexical diversity can and cannot tell you
Lexical diversity is a long-studied measure. It originates in mid-century work on child language development, where a widening ratio tracked growing vocabulary. It correlates with education and with reading exposure.
What it does not do is measure intelligence, and the axis name overclaims. A precise, plain-spoken explanation using few words well is one of the harder things to do in public speaking, and this measure marks it down. A speaker padding with synonyms scores higher.
We use it because it is the only signal on this axis that is not a model's opinion, and an axis fed by one opaque number would be worse. That is a real reason, though it does not make the measure right.
Six ways the number goes wrong
Length dependence, as above. This is the single biggest problem with the number.
It counts words, not ideas. A wide vocabulary deployed to say nothing scores well.
Technical speech is penalised twice. A domain expert explaining one concept precisely will repeat the term for that concept, because using a synonym would be wrong.
It is unusually sensitive to transcription quality. A misheard word is, by definition, a new unique word, so transcription errors inflate the ratio.
Two incompatible versions of this measure are plotted on one chart. A recorded practice counts every word of any length; a live session counts only words of three letters or more, and only unaccented ones. The two series sit on the same axis of the same graph under the same label, and a step between them is as likely to be the change of instrument as a change in you.
And the model half is unaudited. Where the ratio is at least checkable, the model's judgement of "argument sophistication" is not something you can appeal or reproduce.
Raising it without gaming it
Specificity, more than vocabulary. The reliable route up is replacing abstractions with the concrete thing you mean: a number, a name, a date, an object. This raises the ratio as a side effect and improves the speaking regardless of the score, which is the only kind of advice worth giving about a measure this rough.
The unreliable route is reaching for longer words, which raises nothing the model reads as depth and costs you on Clarity.
References
- McCarthy, P. M., & Jarvis, S. (2010). MTLD, vocd-D, and HD-D: a validation study of sophisticated approaches to lexical diversity assessment. Behavior Research Methods, 42(2), 381–392. https://doi.org/10.3758/brm.42.2.381