Poker Table Food Chain

Every home game is tangled: some money flows straight uphill from fish to shark, some just churns in circles between three players who keep trading the same stack back and forth. This tool takes a chip-transfer log and mathematically splits it into those two pieces — the clean hierarchy and the leftover chaos.

Read the ladder as "who is structurally upstream of whom," not moral judgment — a player can sit low on the ladder after one unlucky session and it means nothing about next week.

Grounded in "Disentangling Large-Scale Supply Networks: f-HiCoNE Framework for Flow-Hierarchical Clustering via Combinatorial Hodge Decomposition" — arXiv:2604.04538. The real finding: real supply/transaction networks (the paper studies ~650,000 firms) are not clean acyclic hierarchies — they're entangled with cyclic flows. Combinatorial Hodge decomposition splits any weighted directed flow graph into a gradient component (an acyclic hierarchy, described by a scalar potential per node, solved via a graph-Laplacian least-squares system) and a cyclic/curl component (pure circulation that doesn't move anyone's rank). The potentials line up with real upstream/downstream firm positioning. Scale note: this demo runs the identical linear algebra — build the weighted flow graph, solve the graph-Laplacian least-squares system for a scalar potential, subtract to get the residual — on a poker table instead of 650,000 firms, and stops after the Hodge split (it doesn't reproduce the paper's downstream community-clustering stage).

1Feed it a chip-transfer log

One transaction per line (commas also work): Winner beats Loser amount — e.g. Shark beats Fish1 400. Amount is the size of the pot that changed hands.

Players

–

Chips in circulation

–

Hierarchy score (variance explained by rank, not churn)

–

2Watch it split

Raw table graph

Every transaction as it actually happened — winner and loser, arrow points toward whoever received the chips. This is the messy, uninterpreted input.

▼ combinatorial hodge decomposition (graph-Laplacian least squares) ▼

🪜 Hierarchy ladder (gradient)

Players ranked by scalar potential Φ — the acyclic ranking that best explains the flows. Arrows climb the ladder; width = strength of that hierarchical edge.

🌀 Chaos loop (curl / residual)

What's left after the hierarchy is subtracted out — chips circulating in loops that don't move anyone's rank. Thin/absent edges mean that relationship is pure hierarchy, no churn.

3Full breakdown per player

Rank	Player	Potential Φ	Net winnings	Total volume

What's actually being computed.

1. For every pair of players, net their transactions into one signed flow x_ij (net chips moved i→j) and one confidence weight w_ij (total volume traded between them).

2. Build the weighted graph Laplacian L from w_ij, and the divergence vector div(x). Solve (L + J/n)·Φ = -div(x) for the scalar potential Φ — a small linear system, solved here with real Gaussian elimination (no library). J/n is the standard trick for pulling the minimum-norm solution out of a graph Laplacian's singular null space (potentials are only meaningful up to a shared constant, so we recenter them to mean zero).

3. The gradient component on edge i→j is Φ_j − Φ_i — the flow a clean hierarchy would predict. The residual x_ij − (Φ_j − Φ_i) is the curl/cyclic component — the part no ranking can explain, because it's circulating.

4. Hierarchy score = 1 − (weighted residual sum of squares) / (weighted total sum of squares). 1.0 = perfectly hierarchical, 0.0 = pure circulation (see the "pure circular churn" preset, where it's exactly 0).