spb/wpai0001_uqo
Public
1<!-- Author: Simon-Pierre Boucher — simon-pierre.boucher@uqo.ca -->2# 🧠 The Neural Organ Transplant Experiment34**UQO Working Paper No. AI-0001** — *Causal Transfer of Cognitive Capabilities Across Local Large Language Models*56[](paper/main.pdf)7[](pyproject.toml)8[](tests/)9[](docs/REPRODUCIBILITY.md)10[](https://uqo.ca)11[](mailto:simon-pierre.boucher@uqo.ca)1213> **TL;DR** — Can a bounded, *causally identified* subsystem that implements a measurable capability in a14> donor language model be transplanted into a recipient that demonstrably lacks it, without full-model15> training? This repository holds the preregistered protocol, the experiment harness (toy transformers with16> exact hook semantics; MLX/Transformers backends planned), the registry of runs and the paper.17> **Result (Tier 0, preregistered confirmatory run, 2026-11-09):** a donor-in-the-loop activation graft —18> the recipient's post-layer-0 residual row at every read position replaced by the donor's row through an19> orthogonal Procrustes map fitted on *task-free* rows — lifts the recipient from chance (0.06–0.10) to 0.96620> on held-out Ω programs (control-adjusted effect +0.875, lower bound above 0.25 in 9/9 pairs; control stratum21> 0.992), carries the rule to unseen operands (0.77 vs 0.94), needs eight of 128 dimensions, survives width,22> depth and stride-2 mismatches, and is anchored by ablation → rescue (0.11 → 1.00; sham 0.11). Regression-fitted23> maps (ridge, CCA) annihilate the capability (Proposition 1); donor-free adapters cache; conventional learning24> on the same items is perfect. Tier 1 (Qwen2.5-1.5B/3B, Llama-3.2-3B) and Tier 2 (Qwen2.5-7B) runs: see status.2526---2728## 📌 Status2930| Item | Value |31|---|---|32| Version | 0.1 protocol draft (November 2026) |33| SSRN | not yet submitted |34| Journal | to be assessed after results (ML / mechanistic interpretability venue) |35| Code & data | synthetic data regenerated from seeds; every run has a manifest in `experiments/manifests/` |36| Gates | A ✔ · B ✔ · C ✔ · D ✔ · E ✔ (Tier 0 confirmatory) · F in progress |3738## 🎯 What this paper does3940Prior work maps whole residual streams or concept directions between models (model stitching, activation41buses, cross-model steering) and the one controlled attempt at activation-mediated *capability* transfer42reports a null. We (i) causally localize a capability in the donor by interchange interventions,43(ii) manufacture or verify the recipient's lack of it, and (iii) compare each transplant family44(activation, subspace, circuit, weight, adapter, persistent) against matched sham, random and45unrelated-donor controls on held-out, leakage-controlled tasks. Estimands:4647```48Δ_T = Q_T(R') − Q_T(R) transfer gain49Δ_T^adj = Δ_T − [Q_T(R_sham) − Q_T(R)] control-adjusted transfer (primary)50ρ_c = Q_c(R') / Q_c(R) retention on control strata51N_D = Q_T(D) − Q_T(D \ S_D) donor necessity52```5354## 🏆 Headline results5556| Object | Estimate (run) |57|---|---|58| Donor vs recipient, target stratum | 1.000 vs 0.092; Ω-step 1.000 vs 0.062; control 1.000 both (`pilot_tier0_opseq_451ef5087182`) |59| Localization of Ω in the donor | residual row at the read position restores 1.00 at every layer; L0 attention 1.00; L0–L2 MLPs 1.00/0.90/0.83; sham 0.00 (pilot) |60| **Primary graft (confirmatory)** | target 0.966 (min 0.889), control 0.992 (min 0.946); Δ_T^adj +0.875, lower bound > 0.25 in 9/9 pairs → H2 supported; H4 (≥ 0.95 retention everywhere) not supported (`graft_tier0_confirmatory_0c2257f6803e`) |61| Controls (confirmatory) | sham map 0.091 (control 0.100), norm-matched random 0.091, shuffled item 0.103, shuffled layer 0.186, unrelated-skill donor 0.076 |62| Subspace k = 8 vs random k = 8 | 0.927 vs 0.077 (+0.850 [+0.837, +0.862]) |63| Rule vs cache | graft 0.941 seen / 0.768 unseen (ratio 0.82, H5 supported); donor-free adapter 0.702 / 0.000 |64| Ablation → rescue | 0.113 → 1.000 (sham 0.106) |65| Conventional learning on the same 3 000 items | supervised adapter 1.000; distillation 1.000 |66| Tier 1 / Tier 2 | runs in progress (`tier1_taskvector_*`) |6768## 🗂 Repository layout6970```71wp-ai-0001/72├── CLAUDE.md · docs/MANDATE.md # binding rules and the verbatim research mandate73├── src/wp_ai_0001/ # datasets (OPSEQ/FSA/MORPH), models + hooks, training, localization,74│ # evaluation, statistics, instrumentation, orchestration, visualization75├── tests/{unit,causal_sanity,...} # generators, bootstrap, hook no-op and intervention algebra76├── scripts/ # audit_environment, run_pilot, build_figures, build_paper77├── experiments/{registry,manifests,summaries}78├── artifacts/ # models, per-item outputs, figures (git-ignored)79├── docs/ # RESEARCH_GAP, PREREGISTRATION, EXPERIMENT_PLAN, COMPUTE_BUDGET,80│ # RISK_REGISTER, REPRODUCIBILITY, EXPERIMENT_LOG, DECISION_LOG, TASKS81├── literature/ # verified arXiv records + BibTeX; unverified candidates listed apart82└── paper/ # main.tex + sections/ + references.bib → main.pdf83```8485## 🚀 Quick start8687```bash88python3 -m pip install -e ".[dev]"89python3 -m pytest tests -q90python3 scripts/audit_environment.py --cluster91python3 scripts/run_pilot.py --quick --device cpu --seeds 0 1 --n-boot 200 # ≈ 25 s CPU92python3 scripts/run_pilot.py --seeds 0 1 2 --steps 3000 # ≈ 5 min on an M5 Max (MPS)93python3 scripts/build_figures.py --run <run_id> && scripts/build_paper.sh94```9596## 🗃 Data9798Synthetic micro-worlds generated from seeds (`src/wp_ai_0001/datasets/`), validated by exact recomputation,99hashed and checked for item-level disjointness across splits; see `data/README.md`. Tier-1 checkpoints100(Qwen2.5-1.5B/3B-Instruct, Llama-3.2-3B-Instruct, bf16) are cached locally and not redistributed.101102## 🔬 Methodology1031041. Gap qualification (Phase 1) — `scripts/run_pilot.py`, `training/train_toy.py`, `evaluation/metrics.py`.1052. Causal localization (Phase 2) — `localization/patching.py` (interchange, ablation, heads).1063. Alignment (Phase 3) — `alignment/` (to implement: Procrustes, ridge, CCA, random, unrelated-data).1074. Transplant families and controls (Phases 4–5) — `transplantation/` (to implement).1085. Robustness, mechanism, confirmatory (Phases 6–8) — planned, see `docs/EXPERIMENT_PLAN.md`.1096. Statistics — `statistics/bootstrap.py` (cluster bootstrap over programs; seed-level summaries).110111## 🖼 Figures & tables inventory112113| Figure | Content | Source run |114|---|---|---|115| `fig_gap_tier0.pdf` | donor/recipient accuracy per stratum, seeds, chance | `pilot_tier0_opseq_451ef5087182` |116| `fig_localization_patching_tier0.pdf` | layer × position restored-answer rate, 3 families | idem |117| `fig_localization_ablation_tier0.pdf` | layer × position accuracy after zero-ablation | idem |118| `fig_localization_heads_tier0.pdf` | layer × head patching at Ω token and read position | idem |119120## ⚠️ Limitations121122- Tier-0 capability is a per-step map once the trace is visible (risk R1); multi-step claims need123 `trace_stride > 1` and Tier ≥ 1 evidence.124- No transplant, alignment or control battery has been executed; the paper contains placeholders.125- Length generalization fails for both roles (absolute positions); it is a stratum, not a target.126127## 📝 Citation128129```bibtex130@techreport{boucher2026organtransplant,131 author = {Boucher, Simon-Pierre},132 title = {The Neural Organ Transplant Experiment: Causal Transfer of Cognitive Capabilities Across Local Large Language Models},133 institution = {Universit\'e du Qu\'ebec en Outaouais, D\'epartement des sciences administratives},134 type = {Working Paper},135 number = {AI-0001},136 year = {2026},137 month = {November}138}139```140141## 👤 Author & contact142143**Simon-Pierre Boucher** — Département des sciences administratives,144Université du Québec en Outaouais (UQO), Gatineau, QC.145📧 [simon-pierre.boucher@uqo.ca](mailto:simon-pierre.boucher@uqo.ca)146147## 🗒 Notes148149- Numbering: project id `wp-ai-0001` (series "AI", number 0001) as requested in the mandate; repository150 `wpai0001_uqo.git` on spbgit, following the FIRC-0001 precedent rather than the `wpN` sequence.151- Deviation from the template layout: `src/wp_ai_0001/training/` added (Tier-0 models are trained here);152 experiment artifacts live under `experiments/` and `artifacts/` per the mandate instead of `results/`.153