ai-safe2-framework

Change Attribution

CLI guide Assessment scope System identity

safe2 evidence attribute compares findings from an explicitly trusted baseline with findings from a proposed revision. It separates inherited, introduced, changed, resolved, and unknown findings without claiming causation or framework conformance.

Why it matters

A current scan alone cannot tell a reviewer whether a finding was introduced by the proposed change. A missing finding is not demonstrably resolved when either assessment had incomplete coverage. This workflow keeps those states separate and binds both sides to the same declared agent system and exact scope artifacts.

Command

safe2 evidence attribute change.json \
  --system-identity system-identity.json \
  --baseline-scope baseline-scope.json \
  --current-scope current-scope.json \
  --output change-attribution.json \
  --strict

The source contains normalized findings for both revisions. A finding_key is the caller’s stable correspondence key. Duplicate keys are rejected. The baseline must declare trusted: true and document its trust_basis. Each scope file is bound by its exact SHA-256 digest.

Interpretation

Status Meaning
inherited The normalized finding is identical on both complete assessments.
introduced It appears only in the current complete assessment.
changed The same finding key exists on both sides, but normalized evidence differs.
resolved It appears only in the baseline complete assessment.
unknown Either side declares incomplete coverage, so direction is not supportable.

--strict writes the result and exits with status 1 when unknown attribution remains. This preserves the evidence needed to diagnose the failure.

Evidence boundary

The command does not run scanners, authenticate the baseline, establish that a change caused a finding, verify deployed contents, prove a control effective, or claim AI SAFE² conformance. Retain provider-native evidence with the source.