Area 4.4

Trace Every Significant Output Back to What Produced It

Six months from now, can we reconstruct why the system said what it said?

Enhanced: 2 2 controls

Who it applies to

Enter for every AI system whose letter is E or A (this area has no Baseline controls). Go as deep as your system letter, and implement in your mode.

Why it exists

Make significant outputs reconstructable: keep the trail from the answer back to the question, the model version, and the sources it drew on. When an output is later challenged — by a customer, a regulator, a court — you must be able to show how it came about, and memory will not serve, because the model may since have changed and the context is long gone. Traceability is what turns “we think it worked like this” into “here is exactly what happened,” and it is the evidentiary backbone of both explainability and accountability.

What to actually do

Controls

Enhanced 2 controls

Control 169

Record, for significant outputs, the input, the model version, and the sources consulted.

EnhancedPreventiveall rolesper system

What it is

Enough to reconstruct how the output arose.

How to implement it · Minimum content
  • Scope: which outputs count as significant (e.g. any output feeding a decision about a person or sent outside)
  • Output identifier, timestamp, and the case or person it relates to
  • Input or prompt as submitted, attachments included
  • Exact model version (the pinned version) and the prompt template or configuration version
  • Sources consulted: document identifiers and passages retrieved
  • Output as delivered, any human edit or approval, and who triggered it
How to start

Define which outputs are significant, switch on logging of these fields for them, and test by reconstructing one past output.

Evidence

The traceability records and a demonstrated reconstruction of a past output.

Mode: light / standard / full
light
Use the tool's own activity log where it captures these fields, and otherwise save input, output and model version to a case folder.
standard
The integration writes a trace record per significant output to a dedicated log owned by the system owner.
full
A central tracing service captures every significant output with full lineage, queryable for audits and challenges.

Control 170

Retain the traceability record for as long as you may have to account for the decision.

EnhancedPreventiveall rolesper system

What it is

Balanced against privacy retention limits.

How to implement it · Minimum rules
  • Retention period per output type, set by how long the decision may be challenged (e.g. claim periods, sector record-keeping rules)
  • Personal data in the trail kept only as far as accountability needs; the rest removed or replaced by references
  • Trails held apart from ordinary operational logs, which follow the shorter privacy limits
  • Protected against alteration and early deletion
  • Deleted automatically when the period ends
How to start

For each type of significant output, ask management or legal advice how long the organisation may have to account for it, and set that as its retention period.

Evidence

The retention policy reconciling accountability and privacy, and the configuration applying it.

Mode: light / standard / full
light
A named person keeps trail folders for the set period in a restricted location and deletes them on a calendar reminder.
standard
The system owner configures separate retention for trace logs and the privacy lead approves the periods.
full
Trails sit in write-once storage with automated retention per output type, reviewed jointly by legal and the privacy function.

Evidence

The traceability records; a demonstrated reconstruction of a past output from them; the retention policy reconciling accountability and privacy.