release/command_ledger.md

Campaign command ledger

All formal scientific work used the inherited command:

uv sync --frozen && uv run python reproduction/reproduce.py --output outputs --seeds 80 && uv run python -m pytest -q reproduction/test_reproduction.py

Formal orchestration:

orx exp run 84a9ea1a-5511-4a48-801c-f34c90db4528 --backend local
orx exp run e0d54283-71a9-4c8f-acdb-87d4684e362a --backend local
orx exp run 9916f66a-6781-49bb-aec1-578799f6a35e --backend local
orx exp run d3e3166c-e1b4-4cb2-91dc-04d2bc6f0ff0 --backend hf --flavor cpu-upgrade --image ghcr.io/astral-sh/uv:python3.12-bookworm-slim --timeout 30m
orx exp run bf6502d5-610a-490b-90a4-1df324525600 --backend local
orx exp run 61529636-153f-4105-ac53-894432124169 --backend hf --flavor cpu-upgrade --image ghcr.io/astral-sh/uv:python3.12-bookworm-slim --timeout 30m

The Claim 4 node had one earlier run that exited 1 before a verdict because a NumPy Boolean could not be JSON serialized. Its scalar-only correction did not change formulas, seeds, thresholds, or gates.

Evidence inspection used orx exp wait, orx runs, and orx logs for each run. Startup used orx projects --json, orx project view, orx exp status, Git status/SHA/branch checks, disk inspection, and environment-variable names only. Paper and verdict retrieval used an explicit User-Agent and SHA-256.

Release checks use:

uv run marimo check --strict notebooks/success_conditioning_tutorial.py
uv run python reproduction/verify_release.py --root . --output outputs/release
uv run python reproduction/blind_traversal.py <fresh-candidate>
git ls-remote origin refs/heads/main

The exact text-only Hugging Face upload paths are in release/upload_allowlist.txt; their hashes are in release/upload_manifest.sha256.