Live badges and verification
The README's five CI badges report actual GitHub checks for main. A green CI badge means
the selected check passed. Pending, unavailable and failed results must remain visible.
The release badge follows the latest stable published release; it is independent of later
documentation changes on main.
| Badge | What it reports | Evidence |
|---|---|---|
| PR CI | Latest push to main: Ruff and focused smoke tests | PR CI runs |
| Integration CI | Current main commit: every required validation job, including Python and browser tests, docs and Docker | Integration runs |
| Smoke Test | Current main commit: all nine Linux, macOS and Windows / Python 3.11–3.13 runs | Smoke runs |
| CI Health | Latest CI Health workflow run on main; open it for the evaluated upstream branch and commit | Watchdog runs |
| Release acceptance | Latest push to main: bounded runtime, real inference, browser, contract, recovery, preservation and packaging checks | Acceptance runs |
| Release | Latest published stable version | Release assets, checksums and evidence |
PR CI and acceptance use GitHub's branch and push-event selectors. Health uses GitHub's native workflow badge for the latest run on main. Its previous check-runs badge combined multiple watchdog invocations on the same commit, retaining an older failure even after a later check passed. The native workflow badge follows the latest result while every earlier run and failure stays available.
Smoke and Integration use Shields' live check-runs endpoint with the exact Smoke matrix and
Integration matrix names and main branch. The matrix results run even after a failure and reject
failure, cancellation or skipped required jobs. Integration includes every validation job and all its
matrix variants; tag-only deployment is outside this main-branch check. These matrix badges retain
the direct check-record source adopted after stale workflow images during the 1.3.0 release.
None of these badges sets a status color or manufactures a passing result. Health also watches PRs;
its jobs execute on the default branch even when their evaluated source branch is a PR. A watchdog
failure can therefore indicate a PR requiring attention. Open the run to see its evaluated commit.
Health runs from distinct events, upstream workflows and source branches use separate concurrency groups. An unrelated PR check can no longer cancel an active scheduled/main health check. The watchdog retains its existing rerun, missing-run detection and branch-protection checks. It requires the intended upstream commit (or current scheduled-run commit), checks the returned SHA and branch, and rejects stale results. A pending run is never a pass, including when a grace or waiting period expires. Automatic reruns do not extend the bounded wait indefinitely. Remediation rechecks the current branch commit before acting, and each run permits at most one automatic retry. An event known to be superseded is left read-only so it cannot restart an older workflow and cancel the current build. Branch-protection verification requires the configured admin token; its absence is reported as a warning, not evidence that repository protection has been verified.
If an image disagrees with a run
- Open the badge and check the branch, event, commit and newest run attempt.
- Inspect cancelled and pending jobs as well as failures. Diagnose the cause before retrying.
- Compare the workflow page with the commit's checks. Badge providers and GitHub's image proxy can cache results, so an image can lag a completed run.
- Allow the image cache to refresh. Do not turn a status green by hard-coding its result, forcing its color, choosing an old passing commit, deleting failures or disabling a check.
The Business and Marketplace demo badges now link to repository integration checks and identify their research scope. The Marketplace's former static “100% coverage” badge links to the validation report instead of claiming an unmeasured percentage. These are repository-level checks; they do not certify a demo as production-ready. License and launch badges are informational links and are not test results.
See capability boundaries and the versioned validation archive in each release for tested behavior, skips and expected failures. Badges are a navigation aid; the commit-specific evidence is the basis for assessing a release.