Reproduce from source¶

Install the open procedural runtime, including the policy overlay, public encoder weights and media requirements. Validate the frozen source before starting a sustained GPU experiment:

python3 scripts/verify_publication.py
export WASMAN_ASSET_PROFILE=open-procedural-v1
.venv/bin/python scripts/run_revision_campaign.py \
  --root artifacts/open-core --stage all --training-workers 1 --evaluation-workers 1
.venv/bin/python scripts/score_revision_campaign.py \
  --root artifacts/open-core --output artifacts/open-core/scores.json

The campaign collects successful, audited demonstrations using fixed candidate streams; trains ACT, DP and BC with seeds 17, 43 and 101; evaluates full ordered 30-reset cohorts; and runs the declared diagnostics. The frozen protocol defines splits, budgets, physical success and checkpoint selection. It requires substantial GPU time and disk space and is not an installation smoke test.

The orchestrator stores receipts, hashes and failed attempts. It skips completed jobs on resume and refuses unexplained partial outputs. Completed raw RGB is archived losslessly after training. Concurrency changes execution scheduling; it does not change the scientific cohorts. Scores may vary across workstation stacks; see execution sensitivity.

Individual stages¶

Use --stage collect-train, --stage evaluate or --stage diagnostics with the same output root and scheduling configuration. For a single model, follow training. Use development states for engineering checks. Do not select checkpoints or tune parameters using the published test resets.

Scope¶

This is a self-contained source workflow. Original precomputed packages remain in the private research archive, as explained in data and models. The procedure preserves source pins and protocols; it does not promise bitwise identical fitted weights or success-rate replication. Historical CAD, HotStab, shipwreck and two-arm studies retain their separate contracts.