Conclusion
The six registered claims are supported by independent mathematics and claim-matched experiments. The evidence combines exact rational calculations with the paper's own model and dataset rather than relying on figures alone.
The real FedDPO sweep used DistilGPT-2 (82M), five domain-disjoint SHP clients
with 90 pairs each, a frozen reference, β=0.1, batch 4, R=10 for the
local-step arm, and R=40,E=2 for participation and staleness. The real
DecDPO sweep used the same model and clients, one local step, a decaying
η_r=2e−5/√r schedule, and four five-node Metropolis graph topologies.
The exact audits cover the theorem quantifiers directly: 432 Theorem 5.1 parameter cells, exact staleness proportionality cells for q=0,1,2,5, and 6,350 Theorem 5.5 lower-bound cells with exhaustive subset compositions. The real controls report their denominators, full parameter-vector variance, gradient norms, losses, consensus errors, topology eigenvalues, and seed spreads on the claim pages.