Decision traceability

Defensibility Element 1 (DE-1): Decision Traceability

Definition

For every AI-assisted decision of consequence, the function maintains a contemporaneous record of:

• Inputs considered (data, prompts, context windows, retrieved documents)
• Options weighed (model outputs, alternative recommendations, human-generated options)
• Human judgment applied (final decision, rationale, approver, and authority basis)

This per-decision traceability record forms the audit substrate for defensible AI use.

Canonical Artefact Composition (Tier 3+ capabilities)

A single DE-1 decision record is the logical combination of:

• Materiality Calibration row (GOV-16)
  • Defines the decision’s materiality, impact horizon, and required assurance level.
  • Links the decision to its Risk Class and Tier classification.
• Evidence Register entry (GOV-13)
  • Captures the concrete evidence used: datasets, documents, model outputs, test results, and validation steps.
  • Records timestamps, provenance, and integrity checks for each evidence item.
• Delegation-Authority Register entry (GOV-14)
  • Documents who was permitted to make or ratify the decision and under what delegated authority.
  • Records any escalation path and approvals for exceptions.

These three artefacts, when joined by a common decision identifier, constitute the per-decision audit record for AI-influenced decisions.

Risk Control Mapping

• Primary control for Risk Class 1 (Hallucination) at the per-decision level.
• Ensures that every AI-influenced decision can be:
  • Traced back to the exact inputs the AI saw.
  • Assessed against the options the AI proposed or influenced.
  • Linked to the human judgment that ratified, modified, or rejected the AI’s suggestion.

This enables post-hoc review of whether hallucinated or unsupported content was present, detected, and appropriately handled.

Tier-Specific Requirements

• Tier 3 capabilities (High-impact, human-in-the-loop)
  • DE-1 must exist for every decision of consequence where AI materially influenced the outcome.
  • The decision record must clearly separate:
    • What the AI produced.
    • What the human accepted, edited, or overruled.
    • The authority basis for the final decision.
• Tier 4 capabilities (Agentic / autonomous)
  • DE-1 additionally incorporates the autonomous-action audit trail as defined in the Agentic Governance Charter (GOV-08), including:
    • Action graph or sequence of autonomous steps taken by the agent.
    • Policies, constraints, and guardrails in force at the time of execution.
    • Triggers, observations, and intermediate states that led to each action.
    • Human interventions (pauses, overrides, approvals) and their timestamps.

This extended traceability is required to reconstruct and justify autonomous behavior that may not be visible in a single user interaction.

Operationalization & Storage

• Location of records
  • DE-1 evidence is typically generated and held in vendor systems (e.g., LLM platforms, workflow tools, case-management systems) as the system of execution.
  • The firm must implement weekly export of all DE-1-relevant artefacts to firm-controlled storage (e.g., internal data lake, regulated archive, or records-management system).
• Retention discipline
  • Minimum retention period is the regulatory limitation period applicable to the decision domain (e.g., financial services, healthcare, employment).
  • Where the capability lifecycle (from initial deployment through decommissioning and post-closure review) is longer than the regulatory limitation period, DE-1 records must be retained for the longer of:
    • Regulatory limitation period, and
    • Full capability lifecycle plus any mandated post-closure review window.

Usage Pattern

• At design time:
  • Map each decision type to its materiality (GOV-16), required evidence (GOV-13), and decision rights (GOV-14).
• At run time:
  • For each AI-influenced decision, automatically generate or update the DE-1 record with:
    • Input snapshot (prompts, context, retrieved data).
    • Model configuration (version, parameters, policy set).
    • Output snapshot (raw and post-processed).
    • Human review and final decision, with authority reference.
• At audit / review time:
  • Retrieve the DE-1 record to:
    • Demonstrate that hallucinations, if present, were detectable and controlled.
    • Show that the final decision was made by an appropriately authorized human (or agent under governed charter) with full visibility into AI contributions.

Framework Positioning

DE-1 is the foundational per-decision control for defensible AI use. It underpins:

• Regulatory defensibility (ability to reconstruct and justify decisions).
• Internal accountability (clear chain of responsibility from AI output to human ratification).
• Continuous improvement (feedback loops from decision outcomes back into model and process tuning).

Without DE-1, higher-order controls (e.g., fairness reviews, performance monitoring, incident response) lack a reliable substrate to attribute outcomes to specific AI inputs, configurations, and human decisions.