J · Reference

Chapter 36. Checklists and Job Aids

One-page references for the activities practitioners repeat.


36.1 Gate entry checklists

G1 — Qualification and Provisional Risk

  • Named capability or stakeholder need, with current and target measure
  • Expected outcome stated as a measure with a baseline
  • Named business sponsor able to stop the work
  • “No AI” alternative considered and rejection recorded
  • Provisional D1–D5 scores present, marked provisional
  • Data path identified, at least in principle
  • No architecture or vendor commitment embedded in the canvas

G2 — Architecture and Data Readiness

  • ADR complete, with model, hosting, grounding, and boundary named in comparable form
  • Alternatives considered and rationale recorded
  • Pattern conformance assessed: conformant, modified with detail, or novel
  • Data Lineage and Sensitivity Record complete, each source named
  • Each grounding source has an owner, a refresh cadence, and a maximum staleness
  • Risk score revisited against now-known architecture, and tier confirmed or changed
  • For procured systems: vendor due-diligence fields complete, or Tier-2 cap or documented acceptance recorded
  • For agentic systems: authority boundary drafted

G3 — Deployment Authorization

  • Technical evaluation complete, pass criteria defined before the run
  • AI Assurance Summary at the depth the tier requires
  • Control Matrix complete, controls mapped to specific risks
  • Named accountable owner recorded
  • Human oversight point confirmed for Tier 2 and above
  • Machine-readable permission boundary in place for agentic systems
  • Kill-switch conditions defined and separately approved at Tier 3–4
  • Any conditions carry a named owner and a date

G4 — Material Change

  • Change classified against the change-type table, with reasoning
  • Classification confirmed by a second person at Tier 3–4 where cosmetic is claimed
  • Delta assurance scoped, with what was and was not re-run recorded
  • Risk score updated where any dimension is touched
  • Post-change tier determined, and approval routed to that tier’s authority

G5 — Retirement

  • Retention or deletion confirmed against the schedule
  • Access revoked, including machine identities
  • Downstream consumers identified from telemetry, not only from documentation
  • Evidence set archived and retrievable
  • Knowledge preservation note recorded

36.2 Risk scoring worksheet

Table 110.

Dimension Question Score Justification
D1 Decision Consequence How bad if wrong, assuming the output is acted on?
D2 Autonomy How much human judgment sits between output and effect?
D3 Reversibility Deficit Can the consequence be undone, not the system state?
D4 Exposure and Scale How many affected, how far, can effects cascade?
D5a Data Sensitivity How sensitive are the inputs?
D5b Model Uncertainty How well understood is the behavior?
D5 composite Higher of D5a and D5b

Impact = average(D1, D4, D5), rounded up = ______ Control Deficit = average(D2, D3), rounded up = ______ Matrix tier = ______ Safety override applies? D1 = 4, or (D2 = 4 and D3 = 4) → floor at Tier 3. Yes / No Final tier = ______ Countersigned by ______ Date ______ Scores adjusted ______

Countersigner’s focus list

  • Any 3 where 4 is arguable, particularly on D1
  • Any D3 justified by rollback capability, not consequence
  • Any D2 of 1 or 2 where D1 is 3 or 4
  • Any matrix result sitting one point from a tier change
  • D4 counting direct users rather than reach of consequence

36.3 Drift triage card

Table 111.

Step Action
1 Classify into one of the eight categories
2 Identify the affected system and its current tier
3 Set severity from the tier-and-category table (Chapter 19, §19.4)
4 Route to the accountable owner for that category
5 Determine endpoint: revert, ratify, except, or escalate
6 Produce the corresponding record; no finding closes without one

Exception: agent boundary divergence is an incident at any tier.

36.4 Change classification quick reference

Table 112.

If the change… Then
Adds a tool, raises a limit, removes a human step, or raises an escalation threshold Major
Changes provider or model family Major
Changes training data Major, full revalidation
Changes user population Major, re-enter deployment authorization
Changes business purpose Major, re-enter at S0
Changes a system instruction substantively Material
Adds a grounding source at or above current sensitivity Material
Updates model version at Tier 3–4 Material
Updates model version at Tier 1–2 Minor
Changes formatting only Minor
Lowers a limit or threshold Minor