spb/wpai0001_uqo

1\section{Limitations, Ethics, and Reproducibility}\label{sec:limitations}23\paragraph{What the micro-world cannot show.} Once the execution trace is visible, the Tier-0 capability is a4per-step map from (operator, operand) to a value; the stride-2 variant forces two operators per prediction but5still a bounded computation. Tier-0 evidence therefore establishes that a \emph{bounded, causally localized6computation} can be grafted; it does not establish that multi-step reasoning or open-ended skills can. The7toy models are tiny (0.80\,M parameters, 4 layers); the location of the organ (layer 0) is a property of8these models, not a prediction for large ones. Donor and recipient share the training distribution except9for $\Omega$; real model pairs differ in far more.1011\paragraph{Donor in the loop.} The graft that works is an inference-time dependency on the donor: both models12run on every prompt. It is a transplant of \emph{computation}, not a saved modification. The donor-free13variants (adapter, weights) are the ones that would matter for deployment and they are the weaker ones.1415\paragraph{Conventional learning wins here.} A supervised adapter with the same parameter count, or16distillation on the same 3\,000 items, reaches perfect accuracy. The comparison is fair (same items) but the17task is easy to learn; the interest of transplantation is the absence of labels and gradient steps, not18efficiency.1920\paragraph{Tier 1.} The open-weight experiments use instruction-tuned checkpoints whose training data are21unknown; ``zero-shot'' accuracy is a property of our prompt frame; the alignment prompts are synthetic;22head-level interventions are unsupported in the MLX backend; three models and four tasks do not support23claims about families or scales in general. Items for two tasks (antonyms, capitals) are drawn from finite24hand-written lists and may be contaminated in pretraining; the invented tasks are not.2526\paragraph{Statistics.} Programs are the bootstrap unit and model pairs the second level; with nine pairs27(three same-initialization), pair-level inference is descriptive. Exploratory sweeps selected the primary28method; the confirmatory run uses new seeds but the same trained models, so model-training variance is not29re-sampled at the confirmatory stage.3031\paragraph{Ethics.} All Tier-0 data are synthetic. Open-weight checkpoints are used under their licenses32(Qwen2.5-1.5B: Apache-2.0; Qwen2.5-3B: Qwen Research License; Llama-3.2-3B: Llama 3.2 Community License); no33derived weights are released. The method moves a computation between models without training and could in34principle be used to move unwanted behaviours; the controls in this paper (sham, unrelated donor) are also the35detection tools for such transfers.3637\paragraph{Reproducibility.} Every run has a manifest with git commit, configuration hash, split hashes, seeds,38library versions and hardware; the registry lists every planned, running, failed and completed run; figures and39tables are regenerated from results files by \url{scripts/build_figures.py} and \url{scripts/build_tables.py};40\url{docs/REPRODUCIBILITY.md} gives the commands (CPU smoke test $\approx$ 25\,s; Tier-0 pilot $\approx$ 5\,min;41Tier-0 transplant run $\approx$ 3\,min on an M5 Max; Tier-1 run minutes to an hour depending on pairs).42