M5_S — GDN-only memory-priority sweep on M5 (SimpleMemVLN)

Each top-level folder is one completed navigation model (S1/, S2/, …). The arms change only how M5's episodic memory ranks old visual K/V rows for its fixed R4/B1200 budget; the priority comes from the Gated DeltaNet layer feeding each full-attention layer. Everything else matches the M5_E1 reference (anhdao69/Streaming_VGGT, folder M5_E1), which is the S0 arm and is not repeated here.

Each folder holds the serving files, README.md (arm definition, training summary, loading), provenance/ (identity, qualification, timing and validation evidence) and SHA256SUMS.json. Folders are added one at a time as training finishes; earlier folders are never modified.

Code: SimpleMemVLN branch m5_gdn_s_sweep, source dabf4ad085d20512640fe0ba3bda72ad4724dc76. Training loss is reported per arm; navigation success has not been evaluated in these folders.

Control: R4B0/

R4B0/ is not an S arm. It is the M5_E1 recipe with the episodic bank removed (R4/B0: only the four most recent observation groups, no priority-ranked older rows), trained the same way to measure what the bank itself contributes. Compare it with M5_E1 (R4/B1200) and with the S arms.