vsa-lm-binder

Final binder run (attempt 4). Research artifact for the adaptive memory layer — not a general language model.

Contents: C_slot.pth (SlotVSA binder arm, step 25000, 31.5M), A_slot.pth (in-context baseline, step 25000, 30.9M), C_slot.log (full training log, 11 live evals on fresh topics: 5/8 at step 2500 -> 8/8 from step 5000), slot_arch.py (exact training + grading code), grading and verification logs, training pools.

Pre-registered gates, all passed: G1 recall 94% at 4 docs and 88% at 8 docs vs 6% for the in-context arm at 8 docs (its 512-token window overflows — the fixed-state point); G2 surgical erase (target gone, co-stored doc intact); G3 empty-binder CE gap 5.66 nats.

Independent re-verification (fresh seeds, separate evaluator code): 20/20 fresh-secret recall including 7/7 random codes; empty binder 0/20; wrong-topic page 0/20 (no fabrication path); asked about topic X while holding topic Y's page it answers Y's secret 9/20 — content-driven reading; 8-page capacity 47/48; surgical erase 8/8; erase-then-rewrite 7/7; channel CE gap reproduced at 5.45 nats.

Limits, measured: 1/8 recall with unseen memo filler, 0/8 with paraphrased questions — reads inside its trained template envelope only. Synthetic 780-topic space, not real documents. Contamination check: 0/20 exact (topic, secret) pairs in the training pool; topics themselves appear in training with different secrets. See the results post at github.com/Rikka3/vsa-lm for full context and honest caveats.

Load: requires the vsa_lm package from Rikka3/vsa-lm. CFG: vocab_size 50257, dim 384, num_heads 6, num_layers 6, max_T 512, chunk_size 32, hybrid_every 3. SlotVSA = VSALanguageModel + SlotLayer (8 pages x 64 slots, zero-init no-op at birth). VSA kernels require sequence lengths divisible by 32 — pad and read logits from the last real position.

Naming: "VSA" is used loosely for the fixed-distributed-state design; the mechanism is decayed linear attention with an editable outer-product state.