Model-version rollback: release weights, preprocessing and calibration together

01 / What belongs to a model version?

At minimum, a version includes weights, preprocessing, label mapping, class descriptions, calibration parameters and thresholds. Replacing only weights while keeping old preprocessing or calibration may still return an output, but it no longer represents the same decision condition.

02 / How does a candidate enter the system?

A candidate first passes development-set regression checks for new and old classes, forgetting, calibration drift, review triggers and runtime cost. Only a passing candidate can be enabled; otherwise the current version remains. A switch records the old version, candidate version, check summary and operator.

03 / Replace “remember to sync” with a manifest

Write the complete configuration as a hashable manifest and check referenced files before release. Rollback then restores a bound set of files instead of searching for an old weight file.

{
  "version": "emotion-gate-2029-03",
  "weights": "sha256:...",
  "preprocess": "sha256:...",
  "labels": "sha256:...",
  "calibration": "sha256:...",
  "checks": ["old-class", "ece", "rollback"]
}

04 / Which failures require rollback?

• New-class performance improves while forgetting old classes exceeds the predefined boundary.
• Calibration error or the risk–coverage relationship degrades materially.
• Hashes for preprocessing, labels, class descriptions or calibration do not match.
• The update times out, the interface fails or the switch cannot restore the complete configuration.

05 / Boundary of the engineering demo

The public React workflow and release-gate examples illustrate states, conflicts and evidence binding. They do not run a real emotion-recognition model or establish continual-learning effectiveness. A real study still needs fixed class orders, replay capacity and independent testing.