Skip to content

Root-Cause Analysis

Root-cause analysis (RCA) explains why an incident may have happened. FDAI stores RCA as a hypothesis with citations, confidence, tier, and evidence check state. It is evidence for a decision, never permission to execute a change.

TierRoleTypical evidence
T0Direct deterministic causeMatched rule, violated control, declared fix
T1Prior-incident reuse or deterministic causal chainResolved incident, ordered change and symptom events, resource dependencies
T2Grounded reasoning for novel or ambiguous casesVouched telemetry, events, rules, knowledge chunks, scenario evidence

T1 reuse re-verifies the prior cause and learned action against current evidence. A T1 causal chain requires a preceding change as its root; a window containing only symptoms holds for review instead of inventing a cause.

If reuse fails re-verification, FDAI does not replay the learned action. It can try a configured T2 reasoner with the current evidence set or hold for human review. Either path records why T1 was rejected, so a similarity hit cannot hide stale scope, changed dependencies, or a superseded fix.

Every citation must come from the evidence set supplied to the reasoner. A malformed response, fabricated citation, unsupported claim, or confidence below the configured threshold becomes an held for review hypothesis and routes to human review.

Telemetry and operator documents are untrusted inputs. Model text cannot override policy, what-if results, or the deterministic verifier.

Confidence comes from verifier, cross-check, and evidence check signals rather than the reasoner’s self-reported confidence. The T2 quality gate requires independent cross-checking, deterministic verification, and citations that resolve inside the supplied evidence allowlist. A rubric or cross-check can only lower eligibility; it cannot rescue an unsupported candidate.

RCA outcomeStored resultResponse path
Grounded and above configured thresholdHypothesis with citationsMay inform a typed proposal
Ambiguous alternativesCapped-confidence hypothesisHuman review
Stale T1 reuseRejected reuse with provenanceCurrent-evidence T2 or human review
Malformed or fabricated citationHeld for review hypothesisNo action; audit and review

A structured T1 chain preserves root and failure event IDs plus ordered hops. Each hop records cause and effect references, lead time, relationship, and confidence. Resource dependency data strengthens related paths and blocks unrelated links when a graph is available.

Temporal order alone is not certainty. Confidence is bounded, reduced when multiple roots explain the failure similarly, and determined by the weakest supported link.

A cited hypothesis still varies enormously in strength. FDAI grades that strength explicitly, so “we think the deployment did it” and “we removed the deployment and the errors stopped” don’t read the same.

GradeWhat it rests onWhat it can support
AssociationThings happened close together on a related resourceExplaining and choosing the next investigation
Predictive precedenceThe candidate repeatedly precedes the effect and predicts its directionA recovery proposal in observation mode or with explicit approval
Quasi-experimentalA comparable untreated group behaved differentlyBounded recovery eligibility when every other safety check passes
InterventionalAn approved intervention reproduced or removed the predicted effectPromotion evidence, and never permission on its own

Two mechanisms keep a grade honest. A chain is scored by its weakest link, so one unsupported hop collapses the whole claim rather than averaging out. And every supporting query is paired with a refutation query: if the hypothesis says errors rose after the deployment, FDAI also asks whether unchanged hosts saw the same rise beforehand. A missing refutation is recorded as unknown, not as support.

When refuting evidence arrives, the hypothesis is marked refuted or dropped to a lower grade and a new revision is written rather than the old one being edited. A decision already made isn’t silently rewritten, but future autonomy can drop if the grade no longer clears the bar policy requires.

Closure is the terminal state after an intervention: confirmed, refuted, inconclusive when the outcome evidence was stale or incomplete, or unsafe when the observed impact went past what was approved. Closure is evidence about the hypothesis. It grants no execution permission, skips no safety check, and a confirmed hypothesis with an unsafe impact moves both the hypothesis and its related action back to observation mode.

Current status: The hypothesis lifecycle, weakest-link scoring, support and refutation links, temporal analysis, and independent closure classification are implemented. The control loop analyzes and audits in observation mode rather than writing causation into the ontology, and a deployment binds its own temporal series, projection publisher, outcome provider, and receipt resolver. No causal result grants execution.

Check these elements together:

  1. Incident and correlation ID.
  2. Tier, outcome, confidence, and evidence check state.
  3. Citations and evidence freshness.
  4. Alternative or ambiguous hypotheses.
  5. Structured causal hops when present.
  6. Linked response plan, decision, mode, and rollback reference.

Missing chain data or evidence renders unavailable. The browser does not reconstruct a more confident explanation than the audit record contains.

To learn aboutRead
How evidence is boundedTriage and investigation
How a mitigation is proposedResponse plans and mitigation
How decisions are auditedRead the audit log
The detailed RCA contractObservability and Detection