spb/wpai0001_uqo

# Paper-specific constraints (binding for every stage of this review)

• Enrolment. WP-AI-0001 was built autonomously on 2026-11-09 (sprints 0–7 in one session) and enrolled the same day for the full pipeline (Stages A–G) under the standing instructions of the earlier waves: Stages C–G run autonomously; every authorization request is treated as authorised EXCEPT paid API spend (0 USD); the review branch is pushed after every commit; main is never touched; one paper at a time; ≤ 7 concurrent subagents.
• Genre. Mechanistic-interpretability / machine-learning methods paper with a preregistered confirmatory test: can a causally localized subsystem ("organ", a metaphor) be transplanted from a donor language model into a recipient that lacks the capability, without training? Tier 0 = toy decoder transformers (0.80 M params, 4 layers) trained from scratch on an invented operator micro-world (OPSEQ, scratchpad trace, modulus 11), donor = all 8 operators, recipient = operator Ω excluded; localization by interchange interventions; six transplant families (activation, subspace, steering, adapter, weight, persistent) against a control battery (sham map, norm-matched random, shuffled item/layer, unrelated-skill donor) and conventional-learning references (supervised adapter, distillation); PREREGISTRATION v1.0 frozen before the confirmatory run (graft_tier0_confirmatory_0c2257f6803e, seed base 90 000); Tier 1 = Qwen2.5-1.5B/3B-Instruct and Llama-3.2-3B-Instruct in bf16 under MLX, task vectors (few-shot − zero-shot last-token state) localized within model and transferred through maps fitted on format-matched task-free prompts; Tier 2 = Qwen2.5-7B pairs on M3U96b. Judge in particular: whether the Tier-0 capability (per-step operator map once the trace is visible) supports the paper's claims; whether "causally identified" is earned (interchange + ablation + rescue + sham); the annihilation proposition (regression-fitted maps remove recipient-unpredictable directions) — its hypotheses, proof and the approximate case; the preregistration (what was frozen, the one failed criterion H4 retention ≥ 0.95, Holm family); the control-adjusted estimand against a sham map that itself destroys the control stratum; same- vs different-initialization pairs (seed = init seed); the donor-in-the-loop nature of the primary graft vs the weaker donor-free variants; rule-vs-cache; the strength of the conventional-learning references; Tier-1 selection on the dev split and the confirmatory split; the MLX backend (no per-head hooks; no Transformers equivalence test); verbs vs descriptive/causal evidence. Referee E: full REFORMS checklist.
• Data and compute. All Tier-0 data are synthetic and regenerated from seeds (split hashes in experiments/manifests/*.json); experiments/registry/runs.jsonl lists every run; experiments/summaries/<run>/results.json + CONFIRMATORY.md; per-item outputs and trained models under artifacts/ are git-ignored (regenerable: scripts/train_models.py ≈ 13 min MPS; scripts/run_transplant.py ≈ 3 min). Tier-1/2 need Apple Silicon + mlx_lm + the bf16 checkpoints (HF cache on the laptop and M3U96a/b; Qwen2.5-7B on M3U96b). Stage B referees do not run inference: lightweight checks on experiments/summaries/ only. Stages D–F may re-run Tier-0 experiments on any node with torch (M4M64a) or the laptop, and MLX inference on M3U96a/b replicas (~/wpai0001_review/), one GPU job per node.
• No paid API call (0 USD). None is needed. Crossref/OpenAlex/arXiv lookups are free and allowed.
• Blinding (addition to COMMON.md). No prior external review exists. docs/MANDATE.md and CLAUDE.md are blinded by COMMON.md (they are the author's instructions); docs/{RESEARCH_GAP,PREREGISTRATION,EXPERIMENT_PLAN,COMPUTE_BUDGET,RISK_REGISTER,REPRODUCIBILITY,EXPERIMENT_LOG,DECISION_LOG,TASKS}.md, literature/, experiments/ are the author's design/audit records: readable.
• Numbers. paper/tables/*.tex are generated by scripts/build_tables.py and paper/figures/*.pdf by scripts/build_figures.py from experiments/summaries/<run>/results.json (paper/figures/FIGURE_PROVENANCE.md names the run per figure); confirmatory decisions by scripts/confirmatory_analysis.py → CONFIRMATORY.md — spot-check the paper against these.
• Single-author paper (Boucher). Never touch the author's checkout ia/wp-ai-0001 nor M3U96b:~/wpai0001.