Skip to content

RCA Evidence Collection Runbook

Use this runbook to assemble a bounded, cited evidence set before accepting or publishing a root-cause analysis (RCA) hypothesis. It separates collection from interpretation so every causal claim can be traced to a known source and time.

Collect only evidence authorized for the incident scope. Store opaque references and hashes instead of secrets or raw restricted payloads.

Begin after incident triage establishes an incident ID, affected scope, accountable owner, evidence time range, and next decision deadline. Repeat collection when the scope or timeline materially changes.

ItemRequired value
Investigation ownerFinal owner person for scope, budget, and final evidence set
ReviewerPerson who verifies source identity, citations, and unsupported gaps
Time boundaryStart and end timestamps, including justified lead-in time
Resource boundaryIncluded resources, dependencies, regions, and explicit exclusions
Evidence budgetSource, query, volume, and retention limits
Access boundaryAuthorized identities and handling requirements

Do not silently widen the collection boundary. Record and approve a revised boundary before querying additional resources or time ranges.

Evidence classWhat to collectReliability checks
Events and auditDetected issues, state transitions, decisions, approvals, actionsProducer, sequence, correlation ID, hash
ChangesDeployments, configuration, catalog, policy, and ownership updatesVersion, actor, scope, completion state
MetricsSLI, saturation, errors, latency, and dependency healthSource, aggregation, missing data, timestamp
Logs and tracesCorrelated execution and request recordsClock, sampling, redaction, trace continuity
InventoryResource state and relationships at relevant timesSnapshot time, source, completeness
KnowledgeRules, runbooks, prior incidents, and architecture referencesVersion, approval state, provenance

Absence of a record is evidence only when the source was expected to emit it and source health proves the absence is meaningful.

  1. Freeze the boundary. Record the incident, target resources, time range, source allowlist, evidence budget, and access scope.
  2. Capture source health. Verify clocks, ingestion delay, retention, sampling, and known gaps before interpreting returned data.
  3. Collect immutable references. Gather correlated events, changes, metrics, logs, traces, inventory, rules, and approved knowledge references.
  4. Normalize time. Preserve source timestamps and record any known clock offset. Do not rewrite timestamps to force an ordering.
  5. Build the chronology. Order supported facts before ranking causes. Mark gaps and conflicting records explicitly.
  6. Form hypotheses. For each candidate cause, state supporting evidence, contradicting evidence, and what observation would disprove it.
  7. Test citations. Verify that every claim resolves to a supplied reference and that the reference was available within the collection boundary.
  8. Review the set. Have the reviewer confirm scope, handling, freshness, citation integrity, alternatives, and unresolved ambiguity.

Use the same fields for the leading hypothesis and its alternatives:

FieldContent
ClaimOne bounded causal statement
Supporting evidenceReferences that increase confidence in the claim
Contradicting evidenceReferences that reduce confidence or support an alternative
FalsifierObservation that would disprove the claim
ConfidenceConfigured confidence value with its basis
GapsMissing or unreliable evidence that affects the conclusion

Do not promote correlation alone to causation. A change near the incident start is a candidate until mechanism and evidence support the causal link.

Stop when evidence exceeds scope, citations cannot be verified, timestamps are inconsistent, source health is unknown, or a provider response may contain unvouched data. Also stop when the collection may expose secrets or restricted payloads beyond the approved handling boundary.

An unsupported result goes to human review as an unresolved hypothesis. It does not become an execution-eligible mitigation.

The final package includes the boundary, source inventory, source-health checks, chronology, evidence references and hashes, hypothesis records, reviewer, confidence basis, and unresolved gaps. Record collection start and end times and link the package version to the incident audit trail.

Collection is complete when every material claim has a verifiable citation, credible alternatives are recorded, handling rules are satisfied, and the reviewer accepts the package. The outcome can be a supported cause, a disproved cause, or an explicit unresolved result.

To continue withRead
Refresh incident scope or severityIncident triage
Route a supported mitigation proposalIncident mitigation and rollback
Preserve the final causal reviewPostmortem workflow