Phase 12 checks (26 to 28 Oct 2026)¶

Evidence for the claims-ledger rows E-70 to E-72. Everything here was run on the development laptop (Windows 11). Nothing was published, uploaded or tagged.

1. Installed wheel, empty folder (E-70)¶

Built at d183423: driftwatch_lifetime-0.1.0-py3-none-any.whl, 206,715 bytes, 89 files; source archive driftwatch_lifetime-0.1.0.tar.gz, 165,857 bytes. scripts/check_release.py --dist and twine check --strict pass on both. The wheel was installed with uv pip install into a fresh Python 3.12 environment and run from a folder that is not a checkout and holds no data:

driftwatch --version printed driftwatch 0.1.0;
driftwatch smoke --out smoke-report printed MATCH: 2 rows, max |diff| 0.000000 (tolerance 0.002) and exited 0.

The capture is docs/architecture/img/term-smoke.png. The same two commands run on fresh Linux and Windows runners in release.yml (job smoke-wheel) and clean-room.yml; those workflows have not run yet because the repository is not public.

2. The cleanup commit changed no stored number (E-71)¶

Baseline taken at f7dffb5, before the edits; compared with the working tree of commit b87c6eb (stale docstrings, one shared file-hash helper, one shared window-key helper, logging instead of print, return annotations):

Check	Result
Smoke run, predictions table	40 rows, 16 columns in both; only `config_hash` and `window_id` differ (both embed the code version), every probability, decision and label is equal
Smoke run, slot table	4 rows, 30 columns; only `config_hash` and `window_id` differ
Smoke run, window table	2 rows, 51 columns; `config_hash`, `window_id` and `fit_seconds` (a timing) differ
`make_tables.py`, `make_e3.py`, `make_e4.py` rebuilt from the stored E1 to E4 predictions	every command exit 0; `git status reports/` empty, so everything these three scripts write under `reports/` (the E2, E3 and E4 tables and figures) equals the committed copy. Not covered: files written by other scripts (`make_e0_table.py`, `make_e1_figure.py`, `label_influence.py`), which were not rerun
`scripts/check_e4_vs_e3.py`	exit 0

3. Filtered public history, rehearsal (E-72)¶

scripts/make_public_history.py --out <scratch> on the repository at 3a5b519: 193 commits cloned as a mirror and rewritten with git filter-repo --path docs/source --invert-paths. Checks: 0 commits touch docs/source; no tree contains it; tags prereg-v1, prereg-v2 and prereg-v3 survive; the number of commits is unchanged (193); the largest blob is 415,219 bytes (data/manifest.json); no match in any revision for the secret patterns the script scans (AWS access key ids, GitHub tokens of the ghp_, gho_, ghu_, ghs_, ghr_ and github_pat_ forms, PyPI tokens, private-key blocks); a pattern scan finds those shapes only and is not a proof that no secret exists. The script printed public history: OK (nothing was pushed) and wrote a map of 193 old and new hashes. The repository before filtering is 13 MB.

4. The deploy step of `docs.yml`, simulated (no ledger row: a behaviour check, not a measured value)¶

The shell of the deploy job was run, unchanged, against a scratch repository with a bare remote, four times: a branch build (creates the gh-pages branch), a tag build (adds v0.1.0/), a later branch build with new content (replaces latest/ and keeps v0.1.0/), and the same content again (nothing to publish). Result: gh-pages holds .nojekyll, index.html, latest/, v0.1.0/, versions.html and versions.json; three commits were made, the fourth run made none. Not simulated: the GitHub token and Pages hosting.

Not checked¶

Anything that needs the public repository: the workflows on GitHub’s runners, the Trusted Publisher registration, a TestPyPI upload, the Zenodo record. These are listed in docs/RELEASING.md with the go each needs.