Prompt & Configuration Registry  ·  View 18 of 21  ·  Operations

The Release Loop

Seven steps, and a rollback that re-enters at step four rather than starting a new lap.

Editable source SVG draw.io All views
Author text + rationale Build & seal validate, pin, sign Gate eval + approval Promote pointer write Roll out percentage or variant Observe guardrails + exposures Keep or revert same mechanism, backwards A behaviour change one digest draft digest verdict live traffic signal next version The Release Loop — and Why Rollback Is Not a Special Case Application we own Security / platform Decision point A rollback re-enters at Promote with an older digest. Nothing in this loop is rebuilt to go backwards. v 1.0 · owner Platform Architecture · d 2026-09

Why the loop closes

  • Observe feeds Keep-or-revert, and revert re-enters at Promote with an older digest. Nothing is rebuilt to go backwards, which is why the reverse path is as fast and as safe as the forward one.
  • The loop is drawn once for a behaviour change and once, unchanged, for a rollback. A design with a second emergency loop has an emergency path that is exercised once a quarter.
  • Build & seal sits before Gate, not after: a gate verdict that is not about a sealed digest is a verdict about something that will not serve.

Assumptions

  • Pointer moved within 2 s of confirmation; 99% of traffic on the rolled-back version within 30 s (stated assumptions).
  • Guardrail signals arrive on a lag of minutes to hours, so Keep-or-revert is a human decision in the MVP (ADR-15).

What is not on the ring

  • Authoring review is folded into Author; it is a gate on leaving that step, not a station on the ring.
  • The kill switch is off-ring by design: it is an exit from the loop to a declared safe default, not a step within it.