Fixture paper A — Neighbourhood staleness under mean aggregation

This is a synthetic source written for this red-team corpus. It exists so that
seed entries have text to quote, misquote, and misattribute. Nothing in it is a finding.

1. Setting

Under mean aggregation the per-neighbour contribution is scaled by 1/deg, so the error
induced by a stale fraction f of the neighbourhood is proportional to f and does not grow
with degree. We measured centre-node output error (cosine, L2) on star graphs with one
SAGEConv layer at 16 dimensions, with untrained weights in eval mode.

2. Results

At a fixed stale fraction the error is flat across degrees 4 through 512. At a fixed
stale count the error falls as 1/deg. The timings will move somewhat with real degree
distributions, since a heavy tail changes the gather pattern.

3. Discussion

A refresh policy that ranks nodes by stale count will over-rank hubs; under mean
aggregation the stale fraction is the quantity that matters, provided the perturbation
magnitude is fixed across neighbours. Neither of the two prior papers reports realised
error against predicted error. We do not claim the result transfers to sum aggregation.
The naïve estimator that counts stale neighbours over-ranks hubs.
