The Four Limitations of Narrative Root Cause Analysis

1. It selects among known hypotheses

A working group lists plausible causes and evaluates each one. The correct cause must already be in the list to be found. Novel failure modes — the ones most worth understanding — are not in the list.

2. It cannot weight competing hypotheses simultaneously

Working groups converge on a conclusion by discussion and consensus. The posterior probability of each hypothesis given all the evidence jointly is never computed. Explaining away — the automatic reduction in competing hypotheses when one is confirmed — does not occur.

3. It has no stopping criterion

Investigation continues until the working group is satisfied, or the deadline passes, or the budget runs out. There is no formal criterion for when enough evidence has been gathered — no equivalent of the posterior exceeding a decision threshold, or the VOI of further investigation falling below its cost.

4. It produces a report, not an updated model

The conclusions of a post-mortem exist in a document. They are not encoded in the model that will be used to make the next similar decision. The next incident starts from the same prior. The institutional knowledge does not compound.

What Changes

A causal diagnostic model addresses all four limitations. Hypotheses are not listed — they are computed as posterior probabilities over every upstream variable in the graph given the observed evidence. Novel causes are represented as unobserved factors (U variables) whose posteriors shift when the standard hypotheses fail to account for the evidence. The stopping criterion is explicit: terminate when the posterior exceeds the decision threshold, or when the VOI of further investigation is less than its cost. And each diagnosis updates the model — the conditional probability tables are revised to reflect what was learned, so the next similar incident starts from a more accurate prior.

The governance implication: a board that reviews post-mortems as narrative reports is reviewing the output of a process that cannot systematically improve. A board that reviews the causal model’s updated parameters is reviewing the improvement itself.

See Diagnostic Reasoning for the full treatment of Bayesian diagnostic inference and sequential VOI.

The Engagement

If your post-mortem process produces reports that sit in a folder — the causal alternative produces a model that gets better every time it is used.

info@rung3.ai