Factory documentation / Help & internals
Help & internals
Answers to common setup questions, followed by the boundaries that make a factory run trustworthy.
Frequently asked questions
Do I need to run every station?
No. Run one station for a focused check; use a Line when ordered composition and Andon rules matter.
Can a failed model run count as clean?
No. Environment authentication failures report unknown findings, not zero; invalid verdicts result in an incomplete run.
Where do the results go?
The default fallback is a local file. A configured Beads destination can receive triaged findings automatically; the target visibility policy controls high/critical routing. See sinks and visibility.
Does a pre-commit hook call a model?
No. Blocking hooks use deterministic work only. Model-based work runs asynchronously.
Under the hood: enumerate, then judge
The pre-pass invokes deterministic tools and produces candidates. An agent skill guides an engine adapter through targeted triage. Discovery and adversarial verification run in separate processes and contexts so the verifier can try to disprove a finding. Empty candidate sets may short-circuit a model call when declared by the station.
Containment is enforced outside the prompt
A declared tier is a ceiling, not an automatic grant. The runner checks capabilities and tool policy. On supported Linux hosts, bubblewrap confines the target read-only and keeps credentials outside the model environment through a broker; unsupported boundaries can cause a refusal or withheld capability. Untrusted-code execution at the highest tier is not available. Review the threat model before changing privilege.
Visibility, severity & noise
Triaged severity drives routing: the configured Beads sink files medium and above, skips low and informational findings, and withholds high and critical when visibility is missing or invalid. Public issues are inputs to triage, not a findings sink. A fingerprint built from station, rule, normalized path and snippet excludes line number; repeat runs dedupe work, while fixed findings can later regress.
- Result A verified finding is eligible for policy-based routing.
- No findings A completed run may be empty.
- Incomplete Missing or invalid verdicts are never reported as clean.