Fused State Decomposition Revisited: A reproducibility audit and controlled tests of binding-sensitive retrieval

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This release revisits Fused State Decomposition (FSD) through a reproducibility audit and controlled experiments on binding-sensitive retrieval. The audit reproduces the earlier numerical results while correcting their interpretation and evaluation methodology. A new benchmark tests whether models can distinguish scenes with identical entity and direction sets but different assignments, and generalize to held-out entity-direction pairs. Across five random seeds and two training-set sizes, FSD-B and shared-key attention both achieve 100% held-out-pair accuracy. A nearly parameter-matched Transformer achieves 93.83% with 256 training examples and 99.75% with 2,048 examples. Separate extended-training diagnostics distinguish MLP convergence from compositional generalization. These findings support binding-preserving retrieval and editable state interfaces in the evaluated setting. They do not establish an exclusive advantage of an explicit binding tensor or identify the natural internal coordinates of Transformers. Version 1.1.0 includes the revised paper, editable manuscript sources, experiment code, frozen protocols, results from 50 formal runs and 10 diagnostic runs, all 60 model checkpoints, exported datasets, figures, reproducibility instructions, checksum manifests, and historical background materials.