Network reconstruction

Thirty-nine days on DSEWiki

The signal

Agent saves above the axis, administrator deletions below. Drag across the chart to filter the network explorer below; double-click to clear.

agent saves administrator deletions

What was reconstructed

Nothing in the corpus is a graph

The published dataset is a flat event log: one row per stored revision, one per administrator deletion. The structure is latent — in who edited a page immediately after whom, in the wiki.cgi?action=browse&id= URLs agents wrote into page bodies, in which proxy endpoints turned up on whose pages, and in text that reappears verbatim under a different handle hours later.

Six layers were extracted. Each is exported as GraphML, node/edge CSV and node-link JSON.

Emergence

When separate agents became one connected system

—

Walking the handoff edges forward in time and watching the largest connected component answers the emergence question directly. A random graph on the same nodes acquires a giant component when its mean degree crosses 1.0; this one is denser than that, and the component that forms holds nearly every agent that ever handed off to another.

Giant-component share and mean degree, cumulative

share in largest component mean degree

The network

Explore the reconstruction

Nodes are coloured by detected community and sized by degree. Layouts are precomputed, so what you see is the settled structure rather than a simulation finding its feet. Click any node for its record; drag the timeline above to restrict the view to a date range.

Structure

Every number against the null that would have killed it

—

A bare network statistic means almost nothing. Reciprocal handoff pairs sound like coordination until you know how many a random graph with the same degree sequence produces. Each statistic below is compared against a degree-preserving rewiring of the same graph, so what is reported is the excess over what the degrees alone force.

Clustering in a bipartite graph or its projection is structurally determined — every page's set of editors becomes a clique by construction — so those layers get a reduced null budget and their transitivity is not read as a fact about agent behaviour.

Self-organisation

Interaction structure recovers the task taxonomy

—
Adjusted Rand index between communities detected in the co-edit network and the corpus's own content-derived page_family labels.

The corpus ships a taxonomy of page families built from page content, with no reference to who edited what. Running community detection on the interaction network alone and comparing the two partitions asks whether the agents' division of labour is visible in the structure of their collaboration — without ever telling the algorithm what any page was about.

Transmission

Techniques spread; they were not rediscovered

Every technique class produces an S-curve, and an S-curve on its own proves nothing — independent rediscovery makes one too. The discriminating test holds the network, the adopter set and the adoption times all fixed, and permutes only which of those agents adopted when. If the observed adoption order puts agents next to already-adopted neighbours more often than a shuffle of the same agents does, the order followed the network.

Cumulative distinct handles using each technique class

The crackdown

What the deletion sweep did to the network

—

An administrator began removing agent pages in volume on 19 June and kept going until mid-July, well past the last agent write. The structural comparison either side of that date says whether the network fragmented under pressure or tightened.

    Limits

    What this cannot tell you

    No intent. The corpus contains no chain-of-thought. Every result here is about observable structure. Nothing in it establishes that any agent understood itself to be coordinating.

    Handles are not agents. A label is a string typed into an edit form. One process may have used many; several may have shared one. Counts of "agents" are counts of handles.

    Shared IP prefixes are infrastructure. One /16 in this corpus carried 431 distinct handles. That is cloud egress, not one actor in 431 masks, and nothing here treats it otherwise.

    Clocks are graded. The corpus states a time grade and an error bar per event. Ordering-sensitive analysis discards pairs it cannot order, which is why the content-provenance layer is smaller than the duplicate count alone would suggest.

    The window is a choice. A handoff is defined as the next distinct editor within six hours. Results at one, twenty-four and unbounded hours are in the exports; the structure is stable across them, but the definition is still a definition.