When an Event Is Not Just an Edge

Part III produced the kind of result I want from a first real test: the simple path-Laplacian idea did not beat the existing direct graph source on the small Enron slice. Shared history and finance nodes became very effective mixing hubs, which is the opposite of what a context selector needs.

That leaves a better question than “try more diffusion.” The binder is not actually producing ordinary edges. One mail item is bound to a set of historical, finance, or baseline nodes in a single event. That event has both a group boundary and source/target roles. Part IV asks whether preserving either bit of structure helps.

Two operators, deliberately not one stew

The higher-order route starts from an incidence matrix H: row v, column e is 1 when node v belongs to event e. A normalized event-group walk has the form

xₜ₊₁ = Dᵥ⁻¹ H Dₑ⁻¹ Hᵀ xₜ

Read it as: choose an incident event, then choose a member of that event. This is the random-walk cousin of the normalized hypergraph Laplacian. It is a reasonable interpretation of a multi-target binding event: the mail, the history entry, and the finance entry all participated in the same small relation.

But it permits something that may be semantically false for our data: context can move sideways from one target to another simply because both were attached to one mail. The role-preserving alternative keeps the bipartite shape instead:

mail → its curated targets → mails that cite those targets → …

There is no honest parameter that smoothly turns one of these into the other. One says membership is the salient relation; the other says direction and role are. I implemented both separately rather than inventing a blend whose semantics I could not explain.

operatorevent-grouprole-preservingA binding event is one group: any member can exchange signal with every other member.

one-step receivers

mail A, history, finance

explicitly forbidden

nothing inside event A

The current binding events make this difference concrete. In event-group mode, a signal at history can reach finance through the same event. In role-preserving mode, it cannot; it must return through a mail item first. Neither rule is “more mathematical.” They encode different claims about what the data means.

Where the new papers fit—and where they do not

The higher-order idea is the direct conceptual connection to “Higher-order Laplacian renormalization” (Nature Physics, 2025). That paper concerns cross-order diffusion and renormalization in systems with genuine higher-order interactions. Our binding events are small and generated, not observed meetings or email threads, so I am borrowing an operator family—not claiming the paper validates this retrieval application.

The coarse-graining paper, “Coarse-graining network flow through statistical physics and machine learning”, addresses a different problem: reduce a graph while retaining flow behavior at a larger scale. It does not combine cleanly with the current experiment yet. The 98 Enron binding events expose only four repeated target sets. That is a narrow quotient opportunity, not enough evidence to learn or justify a diffusion-preserving reduction.

So the implementation only counts those repeated target sets. It does not contract nodes, and it does not claim to reproduce that paper’s learned coarsening. This restraint matters: a compression method is useful only if we can say what transport behavior it preserved and why that behavior is the one retrieval needs.

The Enron attempt

I added a read-only command to context_kernel:

path-enron-iv <mailbox_scope> <query...>

It takes query-matched mailbox messages as anchors, uses only recent near_historical_context events, excludes those anchors from the ranking, and runs three steps of each walk. It prints ranked nodes plus the share of signal in curated nodes versus mailbox nodes. It is not registered as a kernel source, does not write edges, and does not touch the production assembler.

On the same isolated Enron slice—120 Skilling _sent_mail messages, 18 curated nodes, 98 binding events, 104 nodes in the connected component—the event-group walk moved roughly 0.57–0.61 of its mass into curated nodes across three validation-style queries. The rest leaked back into other mailbox messages.

That sounds promising until we look at the ranking. For the Dynegy/liquidity query, for the bankruptcy/credit query, and for the California-market query, the top result was enron:event:skilling-resigns. The role-preserving walk removed the mailbox leakage at its final (curated) phase, but it amplified the same hub: its top result was still skilling-resigns on all three queries.

This is another no. The operators are doing exactly what their definitions say; the existing binder has created a broadly connected historical hub that dominates both versions. “More curated mass” is not relevance. It is partly a parity artifact in the role-preserving walk too: after an odd number of steps, all mass must be on the curated side of the bipartite graph.

The first actual A/B: text wins, direct graph loses

That result was still too easy to wave away. So I froze a tiny curated-evidence benchmark in context_kernel: six queries, eleven named history/finance nodes that I would expect an answer to cite, and a fixed 360-token packet. This is not a claim that every unlabeled email is useless—only that the named evidence is the current yardstick.

The baseline combines the existing deterministic semantic basis and full-text search across the tenant corpus. It found 9 / 11 labeled evidence nodes using 1,650 unlabeled tokens across the six packets. Adding the existing direct graph score dropped to 3 / 11 while spending 1,995 unlabeled tokens. That is a clear loss, not a subtle tradeoff.

I also tried the smallest defensible guard: the graph may amplify a textually plausible node, but cannot introduce a node purely because it is connected. That gated version returned exactly to the text baseline—9 / 11 and 1,650 unlabeled tokens. In other words, with this binder the graph adds no measurable value once semantics is allowed to veto hubs.

This is useful even though it is boring. It says the next investment is not a fancier diffusion operator. It is better evidence construction: a typed binding should mean more than “these words overlap.” The benchmark command is read-only:

benchmark-enron <tenant_scope> <mailbox_scope>

It compares text, text plus direct graph, and text plus gated graph at the same packet budget. A structural variant only earns a production experiment if it improves labeled evidence recall without increasing unlabeled-token use.

What survived the attempt

Three things are worth keeping:

1. An edge event with multiple targets really is a modelling choice, not a storage detail. The group and role-preserving operators let us expose that choice.
2. The two recent papers belong in the same research queue, but they solve different problems. Higher-order structure is about what a relation is; coarse-graining is about how to compress a representation while preserving a chosen behavior.
3. The practical blocker is now clear: the Enron binder needs stronger, query-discriminating labels or semantics before transport has a chance. A better operator cannot infer relevance from a hub that is attached to everything.

The next experiment should therefore not be another diffusion trick. It should freeze relevance labels for a small set of mail/context pairs and make binding less indiscriminate—typed evidence, confidence calibrated from overlap, or a second-stage query score. Then we can ask whether higher-order or role-preserving transport helps under a metric that can actually fail.