KB article
Bidirectional Filtering: Convenience vs Predictability
Bidirectional filters can make models easier to use, but less predictable for AI.
arf-kbcontext-stabilitybidirectional-filteringcontext-volatilityrelationship-path
TL;DR
- Bidirectional filters increase ambiguity.
- Predictability matters more than convenience for AI.
The problem
- Bidirectional filtering allows context to flow in both directions.
- AI queries become harder to reason about.
Why it matters
- Unexpected filter paths create inconsistent answers.
- Explainability suffers when context is unclear.
Symptoms
- Slicer changes affect unrelated metrics.
- AI answers shift when extra dimensions are included.
Root causes
- Bidirectional filtering used to “fix” reports quickly.
- Lack of model governance.
What good looks like
- Directional relationships with documented filter paths.
- Bidirectional filters used only when necessary.
How to fix
- Audit all bidirectional relationships.
- Replace with bridge tables or single‑direction filters.
- Document any remaining bidirectional cases.
Pitfalls
- Leaving bidirectional filters in place because they seem harmless.
- Assuming AI can infer intent.
Checklist
- Bidirectional filters minimized.
- Filter paths documented.
- Context tests pass.
Framework placement
Primary ARF layer: Context Stability. Diagnostic bridge: data-movement-reliability, semantic-reliability, execution-reliability.