Narrative Variability: How a Text Moves Through Its Knowledge Graph


Narrative variability describes how a text moves through its own knowledge graph as it is read. Text network analysis shows what a text holds: its concepts, its topical clusters and the gaps between them. That picture is the same whether you read the text forwards, backwards or with its paragraphs shuffled. Narrative variability adds the order: how far the text travels from its core concepts, how often it changes topic, and whether those movements come in lasting phases or stay even throughout.

InfraNodus measures it by reading the text as a time series and applying the fractal analysis used for heartbeat and breathing rhythms. The result is one name for how the text unfolds — a deep dive, a refrain, chapters or a sweep — and the values behind that name. You can use it to compare interview transcripts before you code them, to see where a draft stops staying with its argument, or to check whether a set of notes has subjects or only fragments. This page is the methodology: what the panel shows, how each series is built, and what every value means.

The Narrative Variability panel in InfraNodus next to a knowledge graph, with a line drawn from the centre of the path to each four-concept window of the text
The Narrative Variability panel. The chart follows the text from start to end, and the lines in the graph show where each part of the text sits in relation to its centre.
Try InfraNodus Free    Jump to the Values
 

What the Panel Shows


Open a graph in InfraNodus, go to Narrative and switch from Topics to Variability. The panel has three parts: a chart of the text from its first statement to its last, the text type, and a table of the values the type is read from.

The chart plots one of three series, chosen with the switch above it. 4-grams follows the path of the text through the graph. Sentences follows the number of words in each sentence. Influence follows how central each concept on the path is. For the 4-grams a second switch chooses between radial (where the text is) and step (how it moves). The left of the chart is the beginning of the text and the right is its end.

The chart and the graph are linked. Move the cursor along the chart and the graph highlights the four concepts the text is on at that moment. Press play and the path is drawn through the network in reading order. In the radial view every window is joined to the centre of the path by a line as long as the value the chart plots: short warm lines are close to the core, long cool ones are far out.

The step view of Narrative Variability: the path of the text drawn from window to window through the knowledge graph and coloured by the length of each step
The step view of the same text. The path runs from window to window: short warm segments stay inside a topic, long cool ones cross to another part of the graph.

Under the chart, the table opens folded. Folded, it lists only the six values that decide the text type, each with a reading of two or three words. Unfold a section, or all of them with the button next to info, to see every value measured for that series, the number of points it rests on, and the control that draws its path in the graph. The copy button next to the panel title puts all the values and their readings on the clipboard as text, ready to paste into Project Notes or a prompt.

The Narrative Variability panel with every section unfolded: 4-grams, sentences, influence, movement and sentence rhythm, each value with a short reading
Every section unfolded. Each row is a value, and a reading of two or three words.

 

How It Works: A Text as a Path Through Its Graph


In an InfraNodus graph the words of a text are the nodes, and words that appear next to each other are connected. Think of the graph as a map: the most connected concepts sit in the centre, and each topical cluster forms a neighbourhood around it. Every concept has a position on that map, so the words of a text, read in order, trace a path across it — sometimes inside one neighbourhood, sometimes across to another.

A single word is a small unit, so the path is read four consecutive concepts at a time. Each such window, a 4-gram, gets one position, the middle of its four concepts, and the window slides forward one concept at a time. A text of a thousand concept mentions gives about a thousand points, in reading order. Two series are taken from them.

Series The value at each point What it follows
Radial The distance of the window from the centre of the whole path Where the text is: near the centre it is on its core concepts, far out it is in detail, examples or digressions
Step The distance from this window to the next one How the text moves: a small step stays in the same neighbourhood of the graph, a large one goes to another
Diagram: one window of four concepts in a graph, its distance from the centre of the path and its distance to the next window, and the two time series built from them
One window of four concepts gives one point in each series.

Two more series come from the text itself. The influence series is the betweenness rank of each concept the text passes through, from 0 (peripheral) to 1 (the main hub); it does not depend on the layout of the graph. The sentence series is the number of words in each sentence. The radial distance has one limit worth keeping in mind: it cannot tell which topic the text is in. A stretch far from the centre may be one topic or several equally distant ones, which is why the step series and the measures of movement between topics are read next to it.


 

How to Read a DFA Exponent


Fractal variability asks whether a text has memory: does what it just did say anything about what it does next? Think of two walkers. The first flips a coin at every step, left or right, with no regard for the step before. The second tends to keep going the way she was going. After a hundred steps the first is still near the start, and the second has made a few long excursions. Their single steps look alike. The difference is in how the steps are organised over time.

Diagram: two walkers over a hundred steps, one without memory and one persistent, and the short and long stretches DFA compares
Single steps look alike. How they are organised over time does not.

The method that measures this is detrended fluctuation analysis (DFA), the analysis used on the intervals between heartbeats to tell an adaptive rhythm from a rigid or a random one. It compares how much a series swings over short stretches and over long ones, and returns one number: the exponent α, which is how fast the swings grow with the length of the stretch.

α The series In a text
around 0.5 Has no memory: each value is independent of the last An even texture, the same everywhere
0.5 to 1 Is persistent: it settles into a state and stays there for a while Phases: a stretch in one place, then a stretch in another
near 1 Is fractal: short episodes nested inside longer ones The same pattern at every scale, from a sentence to a chapter
above 1 Drifts without returning The text leaves and does not come back

The panel reports up to three exponents for a series, fitted over different ranges. α covers all scales: stretches of 4 points up to a quarter of the text. α1 covers the short range, stretches of 4 to 16 points, so it describes what happens from one sentence or step to the next few. α2 covers the long range, from 16 points up to a quarter of the text, so it describes phases. A point is a window for the 4-grams, a sentence for the sentence series and a concept for influence.

This is why the measure is called fractal variability. It is not about how much a text varies: a text can vary a great deal and still be flat, if the variation is the same everywhere. The exponent captures whether the variation is organised — small movements inside larger ones, the way a coastline looks similar from a plane and from a cliff.

Diagram: three lines showing predictable, random and fractal change, with the fractal line keeping its shape when a part of it is enlarged
Fractal change sits between the predictable and the random: its parts resemble the whole.

An exponent is only as good as the series under it. It needs at least 64 points, and the panel says how many each one rests on:

  • Under 64 points: no exponent is fitted, and the section says how many sentences or points the text has.
  • 64 to 127 points: a rough indication only.
  • 128 to 255 points: usable, with noticeable noise.
  • 256 points or more: as reliable as this measure gets for a single text.

 

The Text Type: Four Ways a Text Unfolds


A table of exponents is hard to read, so InfraNodus reduces it to one name. The name answers two yes-or-no questions. Does the text have a core? This comes from the structure of the graph: a biased or focused graph has one topic or a few concepts that carry the text, a diversified or dispersed graph has no single core. Does it stay where it goes? This comes from the time series: a text that stays where it goes (dwells) enters a topic and remains there for a while, and a text that keeps moving (roams) passes through its topics at an even rate.

Diagram: the four text types as strips of colour — deep dive, refrain, chapters and sweep — arranged by whether the text has a core and whether it stays where it goes
Two questions give four types. Dark segments are the core, colours are topics.
Stays where it goes (dwells) Keeps moving (roams)
Has a core Deep dive: goes into its main topic, stays there, comes back to the core and goes out again. A lecture on one subject. Refrain: keeps returning to the same core at an even rate, without settling into any part of it. A pitch that repeats its message.
No single core Chapters: moves through several topics and stays in each for a while. A report with sections. Sweep: touches many topics at an even rate, never settling. An overview or a list.

The answer to the second question is the average of up to six measures, each put on the same scale and weighted equally. Three are DFA exponents: the long-range exponent of the radial series, the short-range exponent of influence, and the overall exponent of sentence length. Three describe the movement between topics directly: how often a step crosses into another topical cluster (counted in reverse, so fewer switches score higher), how uneven the stays inside a topic are, and whether small and large steps come in runs. The last three follow the length of a text, so each is first compared with what is usual for a path of the same length.

Diagram: six measures averaged into one score that says whether a text stays where it goes or keeps moving
Six measures, one answer. These are the six values the folded panel lists.

A measure that cannot be computed is left out, and the remaining ones carry the answer: a text under 64 sentences is judged on five, and a very short one on its movement between topics alone. The info view says how many of the six were measured. When the average is close to the dividing line the type is marked borderline, which means the neighbouring type is nearly as likely. None of the four types is better than the others. A product page should probably be a refrain; a literature review that reads as a sweep may be what it needs to be, or a sign that it never commits to an argument.


 

What Each Value Means


The tables below follow the sections of the panel. Each value is read as low, moderate or high, and the two outer readings are given here together with the band where each starts. A value between the two bands reads as moderate. The bands are a working calibration for graphs built at the default 150 nodes: use them for orientation, not as norms. Values marked type are the six that decide the text type and stay visible when the panel is folded.

4-grams: the path through the graph

The radial values describe the distance of the path from its centre: a high value means the text moves in phases, near the core for a while and then out in the periphery for a while, and a low value means it circles at much the same distance throughout. The step values describe the pattern of the movement, not its amount: a text can switch topic often and still read high, if its small and large steps come in runs.

Value What it measures Reads low Reads high
radial α Memory of the distance from the centre, over all scales The distance changes with little memory: few sustained excursions
below 0.785
Long excursions away from the central theme and back, repeating across scales
above 0.875
radial α2 type The same over the long range: whether phases near the core and away from it last The text circles at much the same distance instead of moving in phases
below 0.63
Long phases: near the core concepts for a while, then out in the periphery for a while
above 0.75
step α Memory of step size, over all scales Small and large steps are mixed with little pattern
below 0.60
Settled stretches of small steps alternate with restless stretches of large ones
above 0.635
step α1 type The same over the short range: whether one step resembles the next few From one step to the next few, step size is close to unpatterned
below 0.67
Neighbouring steps tend to be alike in size: small steps come in runs
above 0.715
step α2 The same over the long range: settled and restless phases No settled or restless phases
below 0.56
Long settled phases and long restless ones
above 0.62
radial spectrum The width of the multifractal spectrum, computed on request: whether one exponent describes the whole series or different parts of the text scale differently One scaling rule holds throughout Calm and turbulent stretches follow different rules

Radial α falls as texts get longer, so compare it only between texts of similar length. Radial α2 does not depend on length, but it can differ between the first and the second half of the same text. The spectrum is available for the sentence and influence series too, and it is reported together with a significance test against shuffled versions of the same series.

Sentences: the rhythm of sentence length

The series is the number of words in each sentence, in reading order. A high value means long and short sentences come in runs: the author settles into a rhythm, stays in it, and nests shorter runs inside longer ones. A low value means lengths may vary, but each one is independent of the ones before it. The series needs 64 sentences.

Value What it measures Reads low Reads high
α type Memory of sentence length, over all scales Sentence length resets often: one length says little about the next
below 0.63
Runs of long sentences and runs of short ones, with shorter runs nested inside longer ones
above 0.745
α1 The same over the short range: 4 to 16 sentences From one sentence to the next few, length is close to unpatterned
below 0.61
Neighbouring sentences tend to be alike in length
above 0.73

Influence: when the core concepts come back

The series is the betweenness rank of each concept on the path: how central that concept is to the graph. Only the short-range exponent is reported. It is a weak signal that can differ between the two halves of a text, so read it as a hint.

Value What it measures Reads low Reads high
α1 type Whether central and peripheral concepts come in runs The core comes back at a steady rate
below 0.565
The core comes back in bursts: stretches on the hub concepts alternate with stretches away from them
above 0.60

Movement: plain statistics of the path

These are not fractal measures. They are statistics of the path the text takes from concept to concept. Distances are in graph radii, the typical distance of a node from the centre of the graph, so they compare across graphs of different size. Distance from the centre says where the text lives on average; its spread says how far it travels out and back.

Value What it measures Reads low Reads high
step The mean distance between consecutive concepts, read against what is usual for a path of the same length Consecutive concepts sit close together: the text stays in one neighbourhood of the graph The text jumps far across the graph between consecutive concepts
topic switches type The share of steps that cross into another topical cluster The text settles inside a topic before moving on
below 35%
The text switches topic often
above 45%
topic stays type The variation in the length of the stays inside a topic About the same time in every topic the text visits
below 0.88
Long stays in some topics, and only a passage through others
above 1.0
from centre The mean distance of the path from the centre of the graph The text is anchored: it keeps coming back to its central concepts
below 0.89
The text lives in the periphery, touching many topics without dwelling on its core
above 0.915
spread The variation of that distance A steady distance from the centre, without going deep into the core or far out
below 0.24
The text travels far out to its edges and comes back to its core
above 0.275

The share of topic switches is shown as measured and rises with text length alone, so compare it between texts of similar length. Inside the text type it is corrected for length first. Mean step is read against what a text of the same length usually shows. Human-written texts tend to sit closer to the centre and to have a wider spread: they return to the core and travel far from it. AI-generated texts tend to sit further out with less spread.

Sentence rhythm: a description of the prose

Plain statistics of sentence length in the prose of the text. Statements under 12 words — headings, list items, captions — are left out, and the section says how many sentences were measured. These values are stable within a text and separate kinds of text, so they describe a writing style. They are not part of the text type.

Value What it measures Reads low Reads high
words/sentence The mean number of words in a sentence Short sentences
below 17.5
Long sentences
above 21
under 8 words The share of sentences shorter than 8 words Almost no short sentences
below 5%
Many short, declarative sentences
above 14%
over 30 words The share of sentences longer than 30 words Few long sentences
below 8.5%
Many sentences are allowed to run long
above 16%

 

Reading the Values Together


No single value says much on its own. Read together, they separate two ways of building a text. One is anchored and varies with memory: it keeps coming back to its core concepts, travels far out into detail, and each phase lasts a while, like a meal served in courses. The other is even: it stays at a steady middle distance, touches many topics in every sentence, and one sentence says little about the next, like the same ingredients blended. Both can cover the same ground. The difference is the structure over time.

Human-written text tends towards the first and AI-generated text towards the second: it tends to sit further from the centre with less spread, and its sentence lengths tend to reset more often. The panel marks with a purple dot the values where that difference held once text length was accounted for: both sentence exponents, influence α1, radial α2, the distance from the centre and its spread.

What the values are not

These are tendencies of moderate strength, not a test of who wrote a text. A set of notes, a list of search queries and a story unfold differently whoever wrote them, and several values change with text length alone. Use the values to describe how a text is built and to decide whether that is what you meant.

Diagram: the same city map walked in two different ways, one district at a time or crossing all districts every half hour
Structure is the map. The narrative is the route.

 

Use It in Your Research


The point of the name is the question it lets you ask: is this how I want the text to unfold, and if not, where does it stop? The answer sits in the same graph you already use for topical clusters and content gaps, so it becomes one more step of the same loop.

  • Interview transcripts and open survey answers. Before you code them in MAXQDA or NVivo, compare how each one unfolds. A respondent who keeps returning to one concern (a refrain or a deep dive) differs from one who covers every question evenly (a sweep), even when both mention the same themes. That difference is a candidate for a new research question.
  • A draft or a literature review. Measure the draft, edit it, measure it again. A text that is a sweep and should be a deep dive needs a core to return to and longer stays in each part. Play the step view and watch for the stretch where the path keeps crossing the graph: select those concepts, inspect the statements behind them, and that is the passage to slow down. A change of order moves the type; a change of vocabulary does not.
  • A knowledge base or an Obsidian vault. A folder of notes that reads as chapters has distinct subjects. One that reads as a sweep is a set of fragments that never stay with anything. The radial view shows which concepts the notes keep coming back to: the candidates for a hub note.
  • Content and SEO work. Compare your page with the pages that rank for the same query as paths, not only as topic lists. If they dwell on each subtopic and yours roams, the content gap is in the sequence.

The same values are available outside the app. The InfraNodus API returns them as the network_dynamics object of every response that includes a graph. The InfraNodus MCP server returns them through the analyze_text_signature tool, with one sentence of interpretation for each value, so an agent in Claude, Claude Code or Cursor can check a text it has written against the type you asked for and explain the result instead of quoting an exponent.


 

Limits


  • A text needs some length. The exponents need 64 points in a series. Shorter texts are judged on fewer measures, and the type is then more often borderline.
  • Length changes the numbers. A longer text takes longer steps and changes topic more often, whoever wrote it. Two texts compare best when their lengths are similar.
  • A text is not uniform. The first and the second half of the same text can read differently. The type describes the whole text, which is why the chart and the playback are there.
  • The graph size matters. The bands on this page are for graphs of 150 nodes. With more nodes the path includes more peripheral words and the values shift, so compare texts at the same setting.
  • It describes, it does not judge. None of the four types is a quality score, and none of these values says who or what wrote a text.

 

FAQ

What is narrative variability?

It describes how a text moves through its own knowledge graph as it is read: how far it travels from its core concepts, how often it changes topic, and whether those movements come in lasting phases or stay even throughout. The graph shows what a text holds together; narrative variability shows how long the text stays with any of it. See how InfraNodus builds the graph.

What does a DFA exponent of a text mean?

It says whether a series has memory. Around 0.5 each value is independent of the last. Between 0.5 and 1 the series is persistent: the text settles into a state and stays there. Near 1 it is fractal, with short episodes nested inside longer ones, and above 1 it drifts without returning. The background is in our article on fractal variability.

What are the four text types?

Deep dive, refrain, chapters and sweep. They answer two questions: does the text have a core, and does it stay where it goes. None of them is a quality score. The first question is the one behind the biased, focused, diversified and dispersed states described in cognitive variability.

How long does a text need to be?

A DFA exponent needs at least 64 points: about 64 concept mentions for the path and 64 sentences for sentence length. From 128 points the value is usable and from 256 it is as reliable as it gets for a single text. A shorter text is still given a type from the measures that could be computed, and the type is more often borderline.

Can narrative variability detect AI-generated text?

No. Human-written text tends to move in phases around a core and AI-generated text tends to stay even, but these are tendencies of moderate strength, and several values change with text length alone. They describe how a text is built, and they are evidence to weigh with other evidence, never a test of authorship.

Where are these values available?

In the app under Narrative, then Variability. In the API as the network_dynamics object, documented in the API reference. In the InfraNodus MCP server through the analyze_text_signature tool, listed with the other MCP tools.


 

See How Your Own Text Unfolds


Paste a transcript, a draft or a set of notes, open Narrative and switch to Variability: you see the text type, the six values behind it, and the path of the text drawn through its topical clusters.


Sign Up     Log In