spb/wpai0001_uqo

1{2 "num": "AI-0001",3 "thesis": "in a controlled micro-world (toy decoder transformers trained from scratch on an invented operator language, donor = all 8 operators, recipient = operator Ω excluded), a causally localized subsystem (the post-layer-0 residual row at the read position, identified by interchange interventions, ablation and rescue) can be transplanted at inference time by a donor-in-the-loop activation graft through an orthogonal Procrustes map fitted on task-free rows, lifting the recipient from chance to 0.966 on held-out Ω programs (preregistered confirmatory run, control-adjusted +0.875, 9/9 pairs above the SESOI, control stratum 0.992, one pair below the 0.95 retention criterion), carrying the rule to unseen operands, needing 8 of 128 dimensions and surviving width/depth/stride-2 mismatches; regression-fitted maps (ridge, CCA) annihilate the capability (Proposition 1), donor-free adapters cache, weight grafts cost retention, steering vectors transfer nothing, and conventional learning on the same items is perfect; on open-weight models (Qwen2.5-1.5B/3B, Llama-3.2-3B; Qwen2.5-7B on a node) task vectors are causally sufficient within a model for some tasks and cross-model transfer is reported as measured under a frozen selection rule",4 "node": "Tier-0 experiments (PyTorch MPS, ≈ 3 min per transplant run, ≈ 13 min for the 17-model matrix) run on the laptop or on M4M64a (torch 2.12 available; `~/wpai0001_review/` replica via rsync of the worktree, never the author's checkout); Tier-1/2 MLX inference on M3U96a/M3U96b replicas (`~/wpai0001_review/`; checkpoints in the HF cache: Qwen2.5-1.5B/3B, Llama-3.2-3B on both, Qwen2.5-7B on M3U96b; mlx_lm 0.31.3), ONE GPU job per node; detached nohup + log + done-marker, pull back only `experiments/summaries/*` and `experiments/manifests/*` (≤ 20 MB) and commit per package",5 "rules": "BINDING (peer_reviews/WP-AI-0001/_constraints.md): 0 USD — no paid LLM/API call (none needed); the frozen PREREGISTRATION text (v1.0 + Tier-1 rule) is never rewritten (dated amendments only in DECISION_LOG); trained models and runs are never deleted (new runs get new ids); never touch the author's checkout ia/wp-ai-0001 nor M3U96b:~/wpai0001",6 "ar": "AR-AI0001-1"7}