Structure: the axis with no instrument
By Noam Shemla · Data & Method
TL;DR Structure has no measurement behind it at all. A language model reads your transcript and scores whether the thing you said had a shape: an opening that earns attention, one clear point, a concrete example, and a decisive close. It is the least instrumented axis and, for most speakers, the one worth the most.
Every other axis is about how you spoke. This one is about what you said, and specifically about whether it was built.
The frame is four parts: a hook, a single point, an example, a close. The model is asked to assess each part separately and to mark it strong, partial or missing.
The frame, in four parts Each of the four is marked strong, partial or missing on every recorded practice. This describes the frame rather than a result. The four carry no fixed weights, and the model returns an overall structure verdict alongside the per-section marks.
A model reading, and nothing else
A language model reading the transcript. No acoustic measurement contributes to this axis.
Measured
- Nothing. There is no instrument for this axis.
Inferred, not measured
- The Structure score, 0–100
- Per-section marks for hook, point, example and close
Why a Live Battle counts for more here
Most sources are trusted less when they are less instrumented. Structure inverts that in one specific place.
The reason is that a timed, real-time, adversarial ninety seconds is a more telling sample of whether someone can build an argument than an untimed solo recording where they could stop and start again. The evidence is of the same technical quality; the situation is more revealing.
It applies to both speakers regardless of who won. Gating it on the result would reward beating a weak opponent rather than performing under pressure, and would compound, with winners' signatures growing faster than their actual skill gap justifies.
The penalty that looks like it belongs to this axis
In a Live Battle, a speaker can buy one extra thirty seconds of speaking time for five XP. Doing so costs five points, on the reasoning that needing more time is itself a structural failure: the frame was supposed to fit.
- +30s maximum extension. One click. That is the whole allowance.
- 5 XP what it costs to buy. Charged to the speaker who clicks.
- −5 points, at judging only. The stored practice session is never rewritten.
The speak-time extension, in full.
Which five points, though? The penalty does something other than what its name says.
Named for one number, subtracted from another The constant is called a structure penalty, and the battle screen shows it as five points off structure. The arithmetic that picks the winner subtracts it from the overall score instead. Those are different subtractions: overall is a weighted whole rather than an average of the five axes, so five points off structure and five points off overall are different quantities. The speaker is shown one and judged by the other.
And it never touches this axis in your Signature. The penalty is applied at battle judging and the stored practice session is left alone, so your Structure score, on the radar and in every trend, is computed as though you had never bought the time.
The memory research behind it
The relevant evidence comes from studies of memory rather than of speaking.
Material presented in an organised hierarchy is recalled dramatically better than the same material presented in a random order, an effect large enough that early experiments reported several-fold differences in recall (Bower and colleagues, 1969).1 Structure is not decoration on top of content. It is most of what determines whether the content survives the walk to the car park.
That is the case for this axis, and it is stronger than the case for any of the acoustic ones. It is also the reason the least measurable axis is the one worth most attention.
Where the model's reading falls short
There is no instrument, so there is no appeal. With pace you can check the arithmetic. Here you cannot; you can only disagree.
It is scored from a transcript, so anything visual is invisible. A talk carried by a diagram will score badly on structure even when the room followed it perfectly.
Short clips have nowhere to put a structure. A thirty-second practice cannot fit a hook, a point, an example and a close without becoming a caricature of one, and the model still assesses all four.
The frame moves with your coaching track, but only on one of the two paths. On the estimated path, choosing a structure replaces the four sections with that structure's own: PREP becomes point, reason, example, point. On the measured path the choice is silently ignored and you are always graded on hook, point, example, close. Which path ran depends on your microphone, not on anything you chose, so the same selection can be honoured or discarded without you being told.
And the definition is ours. Four parts is a defensible frame, widely taught, and not a law of nature. Plenty of excellent speaking does not have this shape.
Building a talk that lands
Deciding the close before the opening. Most weak structure comes from a talk built forward from its first sentence, which therefore stops rather than lands. A missing hook is the rarer cause.
The other reliable move is cutting to one point. The single most common structural failure in practice recordings is three points delivered in the time for one, which the model reads as no point at all, and which an audience experiences the same way.
References
- Bower, G. H., Clark, M. C., Lesgold, A. M., & Winzenz, D. (1969). Hierarchical retrieval schemes in recall of categorized word lists. Journal of Verbal Learning and Verbal Behavior, 8(3), 323–343. https://doi.org/10.1016/s0022-5371(69)80124-6