← ClaudeAtlas

strictnulllisted

Build the control before you compare. Use whenever a claim says a graph, network, connectome, wiring diagram, or interaction map is "structured", "non-random", "more clustered than chance", or "beats a random network" - run degree-preserving null models and report what survives.
tsurutanmen/claude-plugins · ★ 0 · AI & Automation · score 70
Install: claude install-skill tsurutanmen/claude-plugins
# strictnull A graph only looks special against the right control. Most papers compare against an Erdős–Rényi graph (same node count, same edge count). That control cannot tell wiring from degree: a heavy-tailed degree distribution alone produces clustering, short paths, reciprocity, and communities. The claim "the wiring matters" needs a control that keeps every node's degree and rewires everything else. ## When this skill applies - A network statistic is reported as evidence of structure (clustering, motifs, reciprocity, modularity, path length, rich club, assortativity, "small world"). - A model built from a real graph (a connectome as a fixed layer, a knowledge graph as a prior, a social graph as a feature) is said to beat a random one. - Any sentence of the form "X is significantly higher than expected by chance" where "chance" is not defined. ## What to do 1. Install: `pip install strictnull` (needs numpy and igraph; networkx optional). 2. Build the graph from what the user has (edge list CSV, adjacency matrix, networkx graph). 3. Run the comparison with at least 20 draws per control: ```python import strictnull as sn g = sn.Graph.from_csv("edges.csv", directed=True) # or from_edges / from_adjacency / from_networkx rep = sn.compare(g, stats=["clustering", "reciprocity", "avg_path", "assortativity"], n_null=20) print(rep) ``` or from the shell: `strictnull edges.csv --n-null 20 --json report.json` 4. Read three numbers per statistic