See the input → responsibility → output diagram.
The prevention register says, for each project, whether the prevention for a failure that already happened somewhere is in place. It checks three recorded causes and answers present, absent, exempted, not applicable or unknown, with the reason.

Meet the safety warden. Not the same failure twice. Keeps guardrail templates and records which protections each project has, so a known failure is not left to recur. Meet the whole Plumb family.
What it does
Each prevention is a template: a fixed, versioned description of the failure it answers and of what counts as the prevention being in place. The first three come from failures seen in the owner's projects:
- A workflow that starts other builds kept starting them after they failed. Pasta's automation started 264 failing builds on one commit. The prevention is a circuit breaker: the workflow declares that it stops after a set number of failures on one commit.
- Public repositories ran jobs on the owner's own machines (self-hosted runners), where anyone's pull request could run code. The prevention is that a public repository runs no job on a self-hosted runner.
- GitHub turned off a scheduled workflow after a quiet spell, so Tally's production smoke tests stopped giving any signal. The prevention is that no workflow stays turned off for inactivity.
The register reads nothing itself. Another part of the Factory, or a person, reads GitHub and hands over what it read: each project's list of workflows, its workflow files, the jobs of some runs and, where known, whether the repository is public. The register then gives one answer for each project and template.
It never answers present by default. Present needs complete reads. Absent needs a complete read that shows the gap, such as a whole workflow file whose text starts builds with no breaker, from a complete list of workflows. A list cut short leaves the answer unknown even when it shows a gap; what it shows is still listed. A refused request, a file it could not read, a line it cannot tell is a comment or anything it cannot classify also leaves the answer unknown, with the reason. A project with nothing read is unknown for every template.
The owner can exempt a project from a template. An exemption carries the date and the owner's own words, and is shown as exempted, never as present, next to what the reads showed. Each template has a fingerprint (a digest) that changes with any edit to it. An exemption names the fingerprint it was given under, so after a change to the template it is no longer applied and is listed instead.
The register can keep its answers and the owner's exemptions in a command journal. Each one is written once, under a name worked out from what it is about: a set of answers from the moment its reads were taken and the projects it covers, an exemption from its project, template and fingerprint. Saving the same thing again changes nothing and gives back the saved copy, so a save whose reply was lost can simply be repeated. Answers are saved under what they were worked out from, so a repeat still gives back the saved answers after a template has changed, marked as made under the earlier templates. Saving something different under the same name is refused, and the first copy stands. The register saves only answers it worked out itself, never ones handed to it, and checks each saved copy whenever it reads one back.
Whether a repository is public is read through Branch health's read-only access to GitHub, one request for the repository itself. With that read, the self-hosted runner template can answer instead of staying unknown.
How it links
- It uses the same reasons for an unread answer as Branch health, takes the workflow lists and job details Branch health already reads, and reads whether a repository is public through Branch health's read-only GitHub access.
- It keeps its answers and exemptions in the Command journal its caller provides.
- Intelligence declares that it uses no AI model.
Limits and status
Honest limits:
- Answers and exemptions are saved, but nothing shows them on the Dashboard, there is no list of everything saved, and an exemption cannot be withdrawn; only a change to its template stops it applying. Nothing reads GitHub on the register's behalf yet, nothing runs it on a schedule, and no project has been checked live.
- A second, different answer for the same reads is refused rather than replacing the first. Adding an exemption therefore takes fresh reads before the register's answer is saved again.
- No prevention has been added to any project. Adding them (rollouts) is a later step that needs each project's permission.
- The breaker check reads a declaration in the workflow file. It does not show that the breaker works.
- The breaker check only looks for a few known ways of starting builds in each workflow file's own text. A build started from a script, from an action it does not know or in another form is not seen, so "present" and "not applicable" cannot rule one out.
- The breaker check reads each line of a workflow file on its own. It understands the common ways of quoting text on one line in shell, JavaScript and Python, including a quote escaped with a backslash, so a
#inside quotes makes the answer unknown rather than hiding a build start. Quoting that runs over several lines, or an unusual kind of quoting, can still hide one. - Job details cover only some runs, so they can show a self-hosted runner but never that there is none. For a public repository without such a job the answer stays unknown.
- Branch health's regular check does not read whether a repository is public; the separate read must be asked for. Until it is, the runner template is unknown, even when a self-hosted runner was seen. The tests answer that read with made-up replies, labelled as such, so no repository's real visibility is established.
- The tests use one recorded copy of the projects' GitHub data from 29 September 2026. That copy has no workflow files, so the Pasta file in the tests is a labelled stand-in.
- There is no installable package for it yet.
Status: In progress. The first part (working out answers) is integrated; the second part (saving answers and exemptions, and the visibility read) is integrated after independent review and awaits acceptance. Nothing is accepted, rolled out or packaged.
Shared value implementation: common contracts. Owning errors, limits, stored identities and authority remain as specified above.
