Skip to main content
Back to All Six Dimensions
DIMENSION 05Chapters 11–13

Governance

“Who is accountable and can it be stopped?”

Establishes named operational accountability, reconstructible Decision Receipts, and enforceable stop controls.

Three Pillars, One Steady Platform
Three Pillars, One Steady PlatformInstitutional study in named statutory liability, tripartite fiduciary oversight, and verifiable decision receipts.
1. The Core Question

Can any consequential decision be reconstructed, and does a named role have tested power to pause or kill it?

2. The False-Confidence Pattern

Publishing ethical principles or compliance committee minutes is mistaken for operational runtime control.

3. Minimum Production Bar

A control that cannot change runtime system state is not a control.

Interactive Dimension Tool 05Authority Map & Decision Receipt Explorer

Governance: The Six Authority Classes & Immutable Decision Receipts

A control that cannot change system state is not a control. Inspect how governance obligations escalate as requested authority moves from advisory recommendations to autonomous execution.

1. The Six Approved Authority Classes

Select an Authority Level
Mandatory Approver:Senior Underwriter / Clinician
Human Sign-off Boundary:MANDATORY: Human must explicitly accept, modify, or reject
Required Evidence Trace:Full Decision Receipt with point-in-time features & SHAP rationale
Technical Stop Power:Immediate fallback to manual queue

2. Evidence-Grade Decision Receipt

Schema v2.4 (JSON)
{
  "receipt_id": "DQ-REC-20260922-8841A",
  "timestamp": "2026-09-22T14:32:01.402Z",
  "system_id": "SYS-CREDIT-AUTO-04",
  "system_version": "v2.4.1-commit-9f2b1a",
  "authority_class": "Recommendation",
  "consequential_decision": "Commercial Credit Limit Adjustment",
  "input_evidence": {
    "account_id_hash": "sha256:7f83b1657ff1fc53b92dc18148a1d65dfc2d4b1fa3d677284addd200126d9069",
    "features_evaluated": [
      "balance_30d_mean",
      "cashflow_volatility",
      "bureau_score"
    ],
    "features_timestamp_verified": true,
    "data_admissibility_check": "PASSED (No lookahead leakage)"
  },
  "runtime_state": {
    "model_hash": "sha256:e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
    "prompt_version": "sys_underwrite_v4_strict",
    "inference_latency_ms": 38,
    "uncertainty_score": 0.042
  },
  "policy_validator": {
    "status": "PASSED",
    "invariants_checked": [
      "max_credit_delta <= 5000",
      "debt_to_income <= 0.42",
      "sanctions_clear == true"
    ]
  },
  "human_adjudication": {
    "required": true,
    "signoff_role": "Senior Underwriter ID: EMP-4091",
    "signoff_timestamp": "2026-09-22T14:34:12Z",
    "human_action": "APPROVED"
  },
  "final_disposition": "AUTHORITY GRANTED ($2,500 limit increase)",
  "stop_authority": {
    "accountable_owner": "Dr. H. Vance (Head of Credit Algorithmic Governance)",
    "kill_switch_endpoint": "https://governance.internal.bank/kill/SYS-CREDIT-AUTO-04"
  }
}
Diagnostic Evidence Bank

Evidence Conditions for Governance

Evaluate in Full Review
G1Example evidence: Decision owner / approval record

Is there a named decision owner who is accountable for the authority being requested, not merely for the model or platform?

P:HB:HE:HR:HO:H
G2Example evidence: Authority map / action boundary

Is the system authority class explicit, including what it may know, recommend, change or execute, and where human sign-off is required?

P:RB:HE:HR:HO:H
G3Example evidence: Decision Receipt / evidence-grade trace

Can a consequential decision be reconstructed from the evidence, model/runtime state, policy state and action path that produced it?

P:AB:RE:RR:HO:H
G4Example evidence: Stop authority / escalation / tested control

Does a named role have tested authority and a technical mechanism to narrow, pause, withdraw or stop the system when controls fail?

P:AB:RE:RR:HO:H