2.0 KiB
ADR 0002: Glossary configuration and processing readiness
Status: Accepted
Date: 2026-11-10
Context
Operators need to reuse glossary collections across campaigns and select different context and protected terms for each ordered correction. Configuration inspection must remain useful without a transcript, model credentials, or a populated glossary. A processing run must reject an unusable later correction before publishing outputs or spending resources on earlier corrections.
Decision
ConfigKit composes typed, named glossary collections with the application configuration. Each pipeline entry selects context and protection independently; Audita derives consumer views without changing the resolved configuration. For a glossary correction, active recognized spellings are exempt from that entry's protection view. Other protected spellings and validators remain effective.
After configuration resolution, Audita preflight checks the whole pipeline and coordinates output preparation. It consumes the typed configuration and acquired inputs, without reloading configuration or calling a provider. The output session owns preparation, cleanup, and publication. See the configuration and operations contracts for current behavior.
Alternatives considered
A separately parsed glossary per invocation would duplicate composition and make per-entry selection and provenance harder to express. Requiring nonempty collections in configuration validation would prevent inspection and reuse of valid empty definitions. Preparing outputs or checking each entry only when it runs could leave artifacts or incur model work before finding a later error.
Consequences
Configuration and runtime views have distinct responsibilities. Inspection can show configured selections and origins without presenting derived vocabulary as configuration. Preflight and the output session need a clear handoff, and the processing pipeline must carry each entry's own effective settings and identity.