← STACK/HUB · SANCTIAGENT · RUBRIC v1.1

DRIFT/14

Production Drift

Model, prompt, tool, policy, data, and behavior drift — fourteen controls that catch silent behavior change before a customer does.

Method: self-assessment, not legal advice or independent assurance. Answer only where documentary evidence exists.

14 fields · ~3 minutes0 of 14 answered

01. A behavior baseline (eval suite or golden set) existed before the last production change Yes No / unknown

02. Model version changes pass a review + evaluation gate before production Yes No / unknown

03. Prompt and system-message changes are version-controlled with change notes Yes No / unknown

04. Tool, plugin, and MCP server versions are pinned and inventoried Yes No / unknown

05. Agent permission and policy changes are logged with an approver Yes No / unknown

06. Training and retrieval data freshness / representativeness is monitored Yes No / unknown

07. Output drift (format, length, refusal rate, tone) is tracked over time Yes No / unknown

08. Cost and latency drift have monitored thresholds with alerts Yes No / unknown

09. Behavioral anomalies (scope creep, volume spikes, off-hours activity) raise alerts Yes No / unknown

10. Agent changes ship through a canary or staged rollout Yes No / unknown

11. Rollback of a model, prompt, or tool change can be executed inside a defined window Yes No / unknown

12. Drift is reviewed on a schedule (e.g., monthly), not only after an incident Yes No / unknown

13. Upstream provider model updates (silent version swaps) are tracked via changelogs or notice Yes No / unknown

14. Every drift finding has a documented owner and remediation deadline Yes No / unknown

Your drift-control score

Field-by-field

Full Drift Workbook — $79: 14-row evidence register, PSI/canary threshold calculator worksheet, and 90-day remediation schedule. Unlock workbook →