01
Identify the source
Record what system and configuration the change comes from, along with its provenance and any unknown fields.
Output: Source context
How it works
Understand the change. Revalidate what matters.
Six steps connect a changing AI system to the exact controls, evidence and decisions that need attention—without assuming every update demands a complete reassessment.
02
Establish approved state
Use the accepted system and assurance baseline as the reference for comparison.
Output: Governing baseline
03
Detect semantic change
Identify what changed in models, prompts, tools, permissions, data sources or configuration.
Output: Change set
04
Trace dependency impact
Follow supported component-to-control and evaluation relationships; expose unmapped paths.
Output: Affected assurance surface
05
Determine evidence validity
Preserve unaffected proof, reopen affected evaluations and flag uncertainty for review.
Output: Focused revalidation scope
06
Support a governed decision
Present completed work, blockers and reviewer context. Only an explicit authorized decision advances approval.
Output: Decision context
Sample data
Follow one system through a proposed change.
Explore the relationship between its approved baseline, changed components and resulting review scope.
STEERLANE / CREDIT REVIEW ASSISTANT / SYNTHETIC PRODUCT EXAMPLE
System assurance walkthrough
Credit Review Assistant
Baseline v2 compared with candidate v3
Stage 1 of 4: Approved state
Approval creates a reference state.
An approved baseline is distinct from a proposed candidate. The decision has its own review and evidence history.
Approved baseline
v2
Approved reference
Historical decision and bound evidence remain separate from candidate work.
Proposed candidate
v3
Review required
The changed version is not approved merely because it can be evaluated.
Illustrative authored fixture · not live product output or actual approval
Explainable impact
A decision can be traced back to the changed component.
Mapped dependencies give revalidation its scope. An unmapped dependency is a reason for review, not proof that nothing matters.
Assurance dependency path · mapped change
Example path
A semantic AI system change follows the approved dependency graph until Steerlane reaches the evaluations and evidence that may no longer be valid.
Changed component
primary_model
Model configuration changed
Control
model_quality_threshold
Model quality remains required
Evaluation
credit_quality_benchmark
Quality benchmark must be reconsidered
Evidence
evaluation_result
Fresh evaluation result required
Result
The evaluation and supporting evidence are reopened because a known dependency connects them to the changed model component.
See the complete operating workspace.
Switch between sample scenarios, impacted evidence and decision context.