Big Bounce Lab — isolated Hubify reproduction workspace

This is NOT the research repo. The authoritative Big Bounce research is github.com/Hubify-Projects/bigbounce (local sibling checkout; set BIGBOUNCE_SOURCE to its path, default ../bigbounce) and must never be modified through this workspace.

This repository is the isolated reproduction environment used to dogfood Hubify's "Reproduce This Lab" workflow against real research. Protected baseline: git tag safety/pre-hubify-lab-2026-07-23 on the authoritative repo.

What this workspace is

• The reference implementation of a Hubify Lab package (lab.yaml)
• Level 0–2 reproducibility tooling that reads the authoritative repo read-only
• The regression baseline (manifests/, artifacts/) that any future Hubify reproduction of Big Bounce is validated against

Rules

1. Every tool in tools/ opens the authoritative repo read-only. No writes, no script execution inside it, no git operations against it.
2. Divergence between a reproduction and manifests/level0.manifest.json is surfaced, never silently regenerated.
3. Nothing is exported publicly from here until it passes the asset classification in hubify/project-context/bigbounce-flagship/audit-assets-classification.md (private).

Reproducibility levels

See REPRODUCIBILITY.md for the current level status board.

Usage

# Level 0: build the archival-integrity manifest (read-only over the source repo)
python3 tools/build_manifest.py

# Level 1: reconstruct the packages/namaster-proof environment (auto-detects Python >=3.10)
python3 tools/build_environment.py --source /path/to/bigbounce

# Level 2: verify artifacts against the frozen manifest
python3 tools/verify_artifacts.py --root /path/to/bigbounce   # local, full 740-file coverage
python3 tools/verify_artifacts.py --remote                    # no source repo needed, 663/740 subset

# Level 3: re-run the P1B NaMaster window-recovery smoke test (deterministic, <1s, no data/GPU)
python3 tools/smoke_test_namaster_proof.py --source /path/to/bigbounce

See REPRODUCIBILITY.md for exactly what each level's status covers — every "done" is scoped, not a blanket claim.