Preserve creative differences in federated aggregation
Determine whether current federated learning aggregation methods can preserve the differences among contributors in small creative communities rather than smoothing those differences away through averaging.
References
Whether current aggregation methods can preserve that difference rather than smooth it is, to our knowledge, an open question.
Four questions raised by the framework are, to our knowledge, open. (i) Experience gradient: can representation scope, behavioral trust tier, and human-approval gate be unified as a single gradient on one ordered structure, making the authority formula of \S\ref{sec:trust} a theorem rather than an engineering rule? (ii) Order-freedom: commitment semantics are history-free in principle~\citep{singh2000, yolum2002}; if so, temporal weight is an implementation leak rather than an inherent difficulty, and one should be able to prove that sealed answers restore the history-free semantics exactly. (iii) Federated reconciliation: pairwise commitment alignment and epidemic convergence are well studied; convergence whose target is ``agreement plus explicitly preserved disagreement'' among many sovereigns lacks a definition of termination.