DANTE-GraphLENS: Adaptive Slot Discovery of Latent Mechanisms in Relational Graphs

DANTE-GraphLENS extends GraphLENS from epistemic bounding over known characteristics to controlled discovery of latent characteristic slots. The architecture combines a typed relational encoder, competitive slot attention, adaptive cardinality gates, evidence reconstruction, disentanglement pressure, and held-out slot ablations. On a hidden-factor graph benchmark across three matched seeds, Gaussian/ARD gating achieved the strongest adaptive result: in-distribution factor F1 0.642 +/- 0.047, factor-count MAE 0.402 +/- 0.149, and unseen-composition factor F1 0.572 +/- 0.021. Fixed slots recovered nearly all true factors but overpredicted cardinality by 3.45 slots per example. Sparse gates improved cardinality but reduced recovery, while the tested truncated stick-breaking construction was unstable at confirmatory scale. The record includes the 14-page paper and a reproducibility archive containing code, tests, raw per-seed results, aggregate results, and the manuscript builder. The paper explicitly limits its claim to controlled synthetic factor discovery under stated generative assumptions.