The Grounding Seam Failure: useful relations are not discovered from features

Summary

A negative result that sharpens a boundary. We asked whether a world-state model can discover the objects and relations of a world from observations — the "grounding seam" — strongly enough that downstream state quality beats per-object feature models, and survives adversarial review. It does not. The route that looked like a win (a relation-aware graph model beating a feature-only model on a citation graph) is, on inspection, classic label propagation over a label-homophilous graph — not a discovered world-state.

The three boundary findings

1. Feature-derived relations are too noisy to use. Inferring a relation graph from per-object features (a nearest-neighbour construction and several variants) never beats a feature-only model; message-passing over the inferred graph hurts. There is a fundamental tension: a relation recoverable from features carries no information beyond the features (it merely smooths them), while a relation useful beyond features is, by definition, not recoverable from them.

2. Handed homophilous topology helps — but only as label propagation. On a homophilous citation graph, a relation-aware model beats features (roughly 85% vs 75% accuracy). But a decades-old, feature-free label-propagation baseline reaches roughly 85% too — none of the world-state machinery (object encoder, role head, edge reconstruction, closure loss) is load-bearing. The operative variable is the label-homophily of the topology, not any "precision" of a relational channel.

3. On heterophilous recorded channels, no relation model beats features. On real heterophilous graphs (edge-homophily roughly 0.1–0.2), feature-only models dominate (roughly 76–85%); graph convolution, label propagation, spectral-structural, and a heterophily-aware model all fall below them. Where edges do not align with the label, the relational channel adds nothing — sometimes nothing beyond a random graph.

The one-line takeaway

Useful relations are not discovered from features; they must arrive as high-precision events. A learned graph-from-observation model is the wrong shape. The right shape consumes relations that are already observed as high-precision events — invoices, tickets, transactions, calls, calendar entries — not relations guessed from features or text.

What this closes, what stays open

CLOSED: a world-state model framed as learned graph-from-observation (this iteration). STILL ALIVE (untouched by this result): an evidence-governed state runtime that folds observed events into typed state. NOT TESTED: a next-state temporal system; a typed relational-event model. Reopening the learned-grounding route requires a genuinely new mechanism, not a better dataset.