Appendix B. Departures From the Specification
Where this handbook recommends something the specification does not require. Record any you adopt as local variations, so that a future reader can distinguish framework from practice.
Table 114. Departures From the Specification
| Recommendation | Chapter | Rationale |
|---|---|---|
| Instrument override rate for low-D2, high-D1 systems | 6 | Detects effective autonomy exceeding the recorded score |
| Score D5 composite as the higher of its two sub-scores | 6 | Prevents sensitivity being diluted by low uncertainty, or the reverse |
| Bias audit sampling rather than randomize it | 7 | Random sampling of a mostly-correct population wastes the sample |
| Cap open conditions per initiative at three | 5 | Prevents conditional approval becoming unconditional |
| Treat pilot as a deployment state with mandatory expiry | 5 | Closes the pilot exemption route to production |
| Run Control Plane comparison in non-production | 10 | Converts production incidents into development conversations |
| Define maximum acceptable staleness per grounding source | 9 | Makes silent knowledge decay detectable |
| Add aggregate limits for standing-authority agents | 13, 16 | Compensates for the cumulative exposure gap |
| Treat agent boundary divergence as an incident at any tier | 17, 19 | Prevents boundary erosion by case-by-case triage |
| Deploy preventive rules in report-only mode first | 18 | Produces calibration data before enforcement |
| Elevate Security Architect to co-accountable at G3 for agentic systems | 21 | Addresses a limitation the framework itself records |
| Route procurement intake by AI content, not contract value | 21 | Embedded AI arrives in inexpensive software |
| Require a reason when funding concurrence is withheld | 20 | A veto never examined is a veto never questioned |
| Add a funding-side TG0 check | 23 | Converts a detective control into a preventive one |
| Track pattern coverage as a standing metric | 30 | The best proxy for whether the library reduces load |