Reference
Glossary
Sigma — The product and reviewer family. One GitHub App, registered as Sigma Mnemoverse.
Reviewer / lineage — A named reviewer construction (currently Bravo or Charlie). Sibling lineages, not successive versions or a quality ranking.
Version — A generation within a lineage (for example, Charlie v4). Earlier generations keep their own identity and history.
Executor — The recorded stage execution path: gateway uses Sigma's API
keys; owner-login uses a Claude subscription producer with an API verifier;
logins uses Claude and ChatGPT subscription stages. Configuration does not
prove credential readiness, account ownership or the viewer's payer identity.
Producer — The stage that proposes candidate findings, recorded only as
suspicion.
Verifier — A fresh, independent stage that confirms, rejects or reclassifies a producer's candidates. Only a verifier-confirmed finding is "verified."
Finding — A verified, reported issue tied to specific evidence in the change.
Unverified suspicion — A producer's candidate no verifier confirmed; shown separately, and explicitly not counted as a finding.
Severity (P0–P3) — How much a finding matters: P1 is a wrong or impossible build if followed, P2 needs a decision before building, P3 is wording only. A public legend is planned (#112) but not yet published.
Check — The one neutral GitHub check per reviewer per run; its
conclusion is always neutral and it is never a merge gate.
Run — One execution of a reviewer against one exact PR head; has its own stages, findings, cost and a run page.
Recheck — A bounded, targeted re-assessment of one existing finding, triggered by the maintainer's reply in its thread; outcomes are fixed, refuted, deferred, or open/inconclusive.
Deferral — Marking a finding's resolution postponed, without claiming it fixed. Costs no model work.
Scope note — A recorded statement of what a stage could not do (a file it could not read, a dependency it could not install), attached to the run.
Superseded — A run's outcome when a newer head, base or PR state made it no longer describe the current PR.
Rules verdict — A separate deterministic aggregate Approve/Request-changes review, available only with the installation's required evidence, authority and publication path. Comments keeps individual findings without that aggregate. A configured Rules setting is not proof of an enabled or successful live verdict.
Secret guard — The check that withholds any model-authored text shaped like a credential before it is stored or published.