proffesor-for-testing__agentic-qe/docs/implementation/adrs/ADR-035-causal-discovery.md

ADR-035: Causal Discovery for Root Cause Analysis

Field	Value
Decision ID	ADR-035
Status	Accepted
Date	2026-01-10
Author	Architecture Team
Review Cadence	6 months

---

WH(Y) Decision Statement

In the context of v3 AQE root cause analysis for test failures, flaky tests, and quality regressions across multi-layer test execution,

facing correlation-based analysis that conflates correlation with causation, manual investigation effort requiring human expertise, reactive analysis only after failures occur, and no automated guidance on optimal intervention points,

we decided for causal discovery using spike-timing correlation (STDP) to automatically learn causal relationships between test events for root cause analysis and intervention planning,

and neglected correlation-based analysis (conflates correlation with causation), manual root cause analysis (requires human expertise), and statistical Granger causality (requires more data, higher latency),

to achieve automated causal inference from event streams, root cause identification accuracy >70%, optimal intervention point suggestions, and proactive failure prediction,

accepting that causal learning requires sufficient observation time, STDP thresholds need tuning per project, and confidence depends on event observation volume.

---

Context

Current v3 AQE root cause analysis is correlation-based: log pattern matching, temporal coincidence detection, and manual investigation. This approach conflates correlation with causation, requires human expertise, is purely reactive, and provides no intervention guidance.

RuVector's Causal Discovery SNN demonstrates a powerful pattern using spike-timing cross-correlation with asymmetric temporal windows that naturally encodes Granger-like causality. After learning, the weight W_AB reflects causal strength from A to B. If event A consistently precedes event B within a time window, the causal weight strengthens.

This creates automated causal inference where the system learns what causes what, enabling root cause identification and optimal intervention point discovery.

---

Options Considered

Option 1: STDP-Based Causal Discovery (Selected)

Uses spike-timing dependent plasticity rules to learn causal weights from event timing. Asymmetric temporal windows encode directionality: events that precede others strengthen causal weights.

Pros:

• No explicit training required (learns from event stream)
• Naturally handles temporal relationships
• Provides intervention point suggestions

Cons:

• Requires sufficient observation time
• Threshold tuning may be needed per project

Option 2: Correlation-Based Analysis (Rejected)

Use statistical correlation between event occurrences.

Why rejected: Conflates correlation with causation; can't determine directionality.

Option 3: Statistical Granger Causality (Rejected)

Use Granger causality tests on time series data.

Why rejected: Requires more data, higher computational cost, and longer analysis time.

---

Dependencies

Relationship	ADR ID	Title	Notes
Part Of	MADR-001	V3 Implementation Initiative	Defect intelligence
Relates To	ADR-034	Neural Topology Optimizer	Shared learning patterns
Relates To	ADR-033	Early Exit Testing	Uses quality signals

---

References

Ref ID	Title	Type	Location
SPEC-035-A	Causal Weight Matrix	Technical Spec	specs/SPEC-035-A-causal-weight-matrix.md
SPEC-035-B	Causal Graph	Technical Spec	specs/SPEC-035-B-causal-graph.md
SPEC-035-C	Discovery Engine	Implementation Guide	specs/SPEC-035-C-discovery-engine.md

---

Governance

Review Board	Date	Outcome	Next Review
Architecture Team	2026-01-10	Accepted	2026-07-10

---

Status History

Status	Date	Notes
Proposed	2026-01-10	Initial creation from RuVector analysis
Accepted	2026-01-10	Approved by Architecture Team