Staged Change

A consolidation computed against a known base and measured before it is applied — what would change, whether the system is better, and against which store it is valid.

Why it matters in a memory layer

Consolidation writes into the source of truth, so "try it and see" costs the thing you were trying to protect. Staging separates the questions a job that writes directly answers all at once and reports on none of.

Two details carry the weight. The retirement set is read off derived_from rather than specified separately, so the change and its provenance cannot drift apart. And the preview must run every step the application runs: on this course's corpus, omitting the scoring pass from the preview alone is the difference between a change that costs five tokens of headroom and one where the exam never passes at any budget — the derived beliefs land unscored and invisible while their eight sources have already been retired.

The fix is structural rather than careful: promotion calls the preview.

Connections

Taught in: Promotion as a Release

Do not confuse with: Snapshot Isolation