One folder per week
Every week lives at its own permanent URL under
/weekN/ with its own page, styles, and generated
data. Nothing gets overwritten when the next week ships.
The weeks
05Language I · tokens, concordances & lexicons latest
Friends, foes & family: what does a link actually say?
Read the sentence behind all 1,784 Marvel links, type each one with an editable lexicon, and test whether enemies cross Louvain borders — then be the ground truth for 60 of them yourself.
Family links stay home under every setting; “foes cross borders” vanishes once the relationship word must sit near the name.
Communities & weighted ties
Does weight rewrite philosophical history?
Follow 1,374 philosophers between unweighted and weighted Louvain maps, inspect the people who cross borders, and test the story against seeds and shuffled weights.
Real strong ties move community borders, but preserve the topological map better than arbitrary placements of the same weights.
Centrality & brokerage latest
Who matters after fame is held fixed?
Enter the Broker Bureau: switch from raw betweenness to a degree-preserving surprise score, open character dossiers, and stress-test the network by removing its central figures.
The most unusual brokers are not necessarily the biggest hubs—or the removals that fracture the network most.
Models & null models
Can a fake network fool you?
Four networks, three of them models, names erased. Play the Null Model Detective, shuffle Marvel while every degree stays fixed, and watch the friendship paradox pile up one draw at a time.
Clustering falls from 0.32 to 0.155 under a degree-preserving shuffle — still 18σ from chance, so hubs explain only half the gap.
Networks, degrees & heavy tails
Who talks about whom?
The first map of 303 Marvel articles. Explore the degree distribution on linear and log–log axes, switch on logarithmic binning and a power-law fit, and roam the full network node by node.
In-degree and out-degree correlate at only r = 0.50: fame and narration turn out to be two different roles.
How this site works
Data is generated, not hand-edited
Each week has a script at the repository root —
analyse_week1.py through analyse_week5.py —
that reads the frozen snapshot in data/ and writes
the JSON its page consumes.
Numbers are asserted
The scripts assert n = 303, directed m = 1,784, undirected m = 1,434, and 17 isolates before writing anything, so a broken snapshot fails loudly instead of quietly.
Run it yourself
uv run analyse_week5.py regenerates the latest data, and
serving the docs/ folder over HTTP reproduces this
site exactly as published.