REPRODUCE.md — WP31 PRISME
All commands run from the repository root with the project venv (uv venv --python 3.12 .venv && uv pip install -p .venv/bin/python -r requirements-lock.txt).
| Step | Command | Expected |
|---|---|---|
| Environment audit | .venv/bin/python scripts/environment_audit.py --smoke |
hardware/device_inventory.json, CPU (and MPS/CUDA if present) forward/backward OK |
| Unit + property + integration tests | .venv/bin/python -m pytest tests -q |
23 passed (15 mandated invariants, controller target-freedom, dense ablation, 5 cache-equivalence tests) |
| Exact parameter counts | .venv/bin/python scripts/count_params.py --all-configs --device meta |
backbone 99,502,848 / 446,744,064 / 1,149,339,648 (= proposal); workspace 2,631,173 / 10,505,221 / 21,655,813 |
| Analytic ops table | .venv/bin/python scripts/estimate_flops.py --config configs/prisme_459m.yaml --T 4096 |
1.296 / 1.666 / 1.787 / 2.219 / 5.633 ×10⁹ ops/token |
| Synthetic data | .venv/bin/python -c "import sys; sys.path.insert(0,'.'); from data.generators.synthetic import write_splits; print(write_splits('data/splits/synthetic_v1'))" |
11 split files + manifest (sha256 as in data/splits/synthetic_v1/manifest.json) |
| Web corpus (optional, ~10 min, network) | .venv/bin/python data/processors/prepare_corpus.py |
290M train / 2.9M val / 2.9M locked test tokens, tokenizer 32,768 |
| Miniature end-to-end pipeline (CPU, ~minutes) | scripts/reproduce_pilot.sh |
tiny backbone → editor bootstrap → labels → controller → routing JSON under experiments/runs/ and results/raw/ |
| Dry runs of every trainer | .venv/bin/python -m training.pretrain --config configs/experiments/pretrain_103m.yaml --dry-run etc. |
exit 0 |
| Tables and figures | .venv/bin/python scripts/make_tables.py && .venv/bin/python scripts/reproduce_figures.py --results results/raw/ |
paper/tables/*.tex, results/figures/*.pdf |
| Claim audit | .venv/bin/python scripts/verify_manuscript_claims.py |
AUDIT PASS |
| Manuscript | cd paper && latexmk -pdf main.tex |
paper/main.pdf |
Full-scale runs (Stage A 103M on an M3 Ultra ≈ 4 h; seed matrices on several nodes) are launched with
scripts/cluster_setup.sh <node>, scripts/launch_seed_matrix.sh, scripts/run_stage_cd.sh and
scripts/launch_103m.sh; their run ids are listed in reproducibility/run_index.md.