cyberneticlibrary
← library

Identify root cause from disjoint evidence

root-causeworkflowsetup L3★0
cetej/STOPA ↗
What it does

Generate disjoint hypotheses from evidence classes, rank via adversarial panel

Best for

Diagnose a failure when multiple evidence sources exist and hypothesis generation needs to avoid anchoring.

Inputs
  • · symptom (string)
  • · evidenceClasses array: {name, content}
  • · maxHypotheses (default 6)
Outputs
  • · ranked hypotheses with verified/refuted/needs-info verdict, panel reasoning per hypothesis
Requires
  • · claude-sonnet (hypothesis generation)
  • · claude-haiku (panel verify/refute)
Preconditions
  • · evidence text is read-only (no agent-written fixes)
  • · symptom is observable
  • · evidence classes are disjoint
Failure modes
  • · all hypotheses refuted (symptom unexplained)
  • · top hypothesis still has low confidence
  • · panel deadlock (all agents agree)
Trust signals
  • · disjoint generation (one agent per evidence class, blind to others)
  • · verify+refute panel consensus
  • · hypothesis ranked by refute robustness