Two-thirds of 302 root cause analyses proposed no fix. The tracker is the control, not the writeup

The arithmetic inside the 106 is the part worth an hour. Those reviews that recommended anything recommended 731 things in total ^[5], which works out to about 6.9 items each ^[9]. That is not a gentle slope from thorough to thin. An investigation either ended with a pile of action items or it ended with an empty recommendation section, and 196 of the 302 ended empty ^[3], roughly 65 percent of the set ^[4].

Then look at what the 731 were made of. Training at 20 percent, process change at 19.6 percent and policy reinforcement at 15.2 percent ^[6] come to 54.8 percent of the total ^[7], or about 401 individual items ^[8]. Kellogg's team judged that the most frequently recommended solutions were the weaker interventions, the ones less likely to prevent recurrence ^[10]. The recurrence duly arrived: multiple event types repeated during the study period despite repeated analyses ^[11].

So there are two failure modes stacked on each other, and they are not the same problem. One is a remedy-strength problem inside the minority of reviews that produced remedies. The other is a much larger completion problem, where a structured investigation into real harm ^[17] ran to the end and named nothing anyone had to do. Review technique addresses neither. The dev.to essay is straight about the limit of the evidence here: nobody has measured that better post-mortems produce worse follow-through, and what the patient-safety literature actually documents is decoupling, with remedy strength varying independently of analysis quality ^[13]. A careful investigation and a bored one land on the same three outputs ^[14].

The essay offers a psychological candidate for why thoroughness might even compete with follow-through, resting on experimental work in which structured ritual lowers anxiety and the neural response to performance errors, and it labels that a hypothesis rather than a finding ^[15]. Treat it that way. The structural reading needs no psychology: if the artifact that leaves the room is a document, the document is what gets produced well, and Peerally and colleagues gave the result a name in a companion paper the same year, organisational forgetting ^[12].

The transfer to software is an analogy the essay draws rather than a count it supplies: the two-hour review, the accurate timeline, and the bug or its sibling back in production inside the quarter ^[18]. If your incident process resembles the medical one in structure, the number to instrument is not review depth. It is the share of reviews that exit with an owned, dated change to a system, and the share of those that are still open ninety days later. Kellogg's institution had eight years of documents ^[1] and repeat harm to go with them ^[11]. Only the tracker, and the engineering work it forces, gets a vote on the next failure ^[16].