Moral Luck and the Postmortem
On the peculiar corporate ritual of grading someone's judgment by consulting the dice they rolled
Two Fridays
Two engineers, two Fridays, one identical decision.
Devi ships a database migration at 4pm on a Friday. She has a rollback plan, a tested script, a canary, and a colleague on standby. The migration works. Nobody notices. On Monday she is a person who ships fast, and her manager mentions her in a channel with a rocket emoji.
Marcus ships a database migration at 4pm on a Friday. He has a rollback plan, a tested script, a canary, and a colleague on standby. A dependency he could not have known about holds a lock, the canary looks healthy for eleven minutes, and then checkout falls over for ninety-four of them. On Monday he is a person who ships recklessly, and there is a meeting about it with a calendar invite that uses the word "learnings."
Same reasoning. Same information. Same care. Different draw from the urn. And the organization, which prides itself on being data-driven, has now produced two entirely opposite readings of the same underlying decision, using a sample size of one, and filed them both under "accountability."
We have a name for this in philosophy, and it is older than your incident tooling.
Bernard Williams Ruins The Retro
In 1976 Bernard Williams and Thomas Nagel, working in parallel and disagreeing productively, put a phrase into the water supply - moral luck. The problem is straightforward and deeply annoying. We like to believe moral judgment applies to what a person controls. Then we notice we do not actually behave that way at all.
Nagel's example is the drunk driver. Two men leave the same party equally impaired and drive home by the same route. One arrives without incident and wakes up with a headache and a private resolution. The other passes a child chasing a ball into the road. We hold the second man responsible for something quite different from the first, though the difference between them was authored entirely by the position of a child in a street.
You can argue that both men were reckless and that only one was unlucky, and you would be right, and this is exactly the point. Our judgments track the outcome far more than the recklessness. The recklessness is priced at roughly nothing until the world provides a corpse.
Then there is Williams' painter, the man who abandons his family to go to Tahiti and paint. If he becomes Gauguin, his choice acquires a retrospective justification that reads as destiny. If he becomes a man with a sunburn and a portfolio nobody wants, he is simply a person who abandoned his family. The moral status of the decision waits, politely, for the market to render a verdict on the paintings.
This is not a puzzle we solved. It is a puzzle we industrialized and gave a recurring meeting slot.
Outcome Bias, Now In Enterprise Form
Psychologists gave the effect a less romantic name. Jonathan Baron and John Hershey ran the experiments in 1988. Give people a description of a physician's decision to operate, hold the reasoning and the probabilities fixed, and vary only whether the patient lived. Subjects rated the identical decision as better when the patient survived, and they also rated the physician's competence higher, and they were more willing to say they would have decided the same way. The knowledge of the outcome bled backwards and repainted the reasoning that preceded it.
Poker players call this resulting, a term Annie Duke has done the most to popularize, and poker players are unusually clear-eyed here because they are forced to be. A player who never confuses a bad beat with a bad bet does not survive a month. The rest of us do not receive the same brutal instruction, because in most organizations you can go an entire career confusing the two, and be promoted for it, on schedule.
Here is the core diagnosis. Expected value is a claim about a distribution. A postmortem is an interview with a single sample. When we treat the sample as a referendum on the distribution, we are not doing accountability. We are doing haruspicy with better slides.
The Blameless Postmortem, A Document In Which No One Is Blamed And Everyone Knows Who Did It
The modern incident review was designed by thoughtful people to solve exactly this, and in many places it works. In many others it has been captured, gently, without anyone deciding to capture it.
You can tell by the tells. The timeline is scrupulously passive-voiced and yet somehow every sentence has an implied subject. The five whys terminate reliably at the most junior person in the room, or at "insufficient testing," which is the organizational equivalent of blaming the weather. The action items are all controls and none are experiments. And the review only ever convenes when something broke, which means the institution has quietly announced that a decision becomes interesting at precisely the moment it becomes unlucky.
Nobody has to say "we punish bad outcomes rather than bad reasoning." The calendar says it. Culture is not what appears in the values deck. Culture is the list of things that summon a meeting.
What This Teaches, Which Is Not What You Wanted To Teach
People are excellent at inferring reward functions from sparse feedback. This is the one machine learning problem humans genuinely excel at. Run outcome-based judgment for two years and you will have trained an organization full of agents who have correctly identified the objective and are now optimizing it with unnerving competence.
The objective is not expected value. The objective is defensibility.
Defensibility and expected value overlap often enough to be confusing, which is why the substitution goes unnoticed. But where they diverge, the trained agent is unambiguous. She picks the expensive vendor with the mediocre product and the immaculate compliance page, because nobody has ever been marched into a hearing for choosing the market leader. He builds the third layer of approval that adds four days and prevents nothing, because the four days are invisible and the prevented nothing is a story he can tell. She declines the experiment with a sixty percent chance of a large win, because forty percent of the time she would have to stand in a room and explain herself, and the win, should it arrive, will be attributed to the strategy, while the loss will be attributed to her.
The most sophisticated form is temporal. The truly fluent defensibility optimizer selects bets whose failure modes mature after his next role change. He is not lying. He is not even cynical. He is simply responding to a gradient you built, and he will be a Vice President before the invoice arrives.
Meanwhile the person who reasoned beautifully and drew badly is quietly recategorized. Nothing is written down. Her name simply acquires a faint asterisk, and she notices, and she adjusts, and the organization loses the one thing it cannot buy - someone willing to make a good bet in public.
The Objection That Deserves A Real Answer
Now for the part that most essays of this genre skip, because it is inconvenient and because "process was sound, outcome was noise" happens to be the favorite sentence of every incompetent person who has ever attended a meeting.
Outcomes are evidence. This is not a bias. This is Bayes. If someone's carefully reasoned bets fail repeatedly, the hypothesis that they are unlucky must eventually yield to the hypothesis that their model is miscalibrated, their probabilities are decorative, and their reasoning is a well-groomed rationalization of what they wanted to do anyway. A person who invokes variance after every loss and mentions skill after every win is not practicing decision science. They are running a small and rather transparent public relations operation.
So the fix is not to stop looking at outcomes. The fix is to stop reading them one at a time.
A single outcome tells you almost nothing about a decision and quite a lot about the world. A pattern of outcomes tells you a great deal about a decision-maker. The unit of judgment for reasoning quality is the portfolio. The unit of judgment for an incident is the system. We currently have these exactly backwards, judging the person by the incident and the system by nothing in particular.
There is a second objection with more weight. Sometimes the outcome is a harmed person, and that person is owed something regardless of how sound the reasoning was. Accountability to those affected is not outcome bias. It is a separate and legitimate obligation, and it deserves its own process, its own remedy, and its own budget. The failure is in fusing the two - in letting the settlement of a harm double as a verdict on a judgment. Repair the harm. Then, separately and coldly, examine whether the decision was wrong or merely unlucky. Blending them produces neither justice nor learning, only theatre with two unhappy audiences.
The Actual Proposal
Review the decision against the information that existed when it was made. Not the information you have now, which you acquired by watching the outcome and which is therefore contaminated beyond use.
Some things that make this real rather than aspirational.
Write it down before you know. A decision record that states the options considered, the reasoning, the probability assigned to the main failure modes, and the conditions under which you would reverse course. Two paragraphs is plenty. Its entire purpose is to be unfalsifiable-proof - a note from a version of you who could not yet know, immune to the retrospective editing your memory will perform without asking permission.
Separate the two reviews explicitly. One asks what the system should do differently. The other asks whether the reasoning was sound given what was knowable. Different questions, different rooms, ideally different weeks. When they share an agenda, the second collapses into the first, and the first collapses into a hunt.
Audit the wins. This is the load-bearing recommendation and the one nobody implements. Run a postmortem on the gamble that paid. The launch that succeeded despite skipping the migration plan. The vendor negotiation won by a director who never mentioned they had no fallback. If the reasoning was poor and the result was good, you have discovered a defect that has been rewarded, promoted, and is currently being modeled for junior staff as best practice. Lucky wins are load-bearing lies. They are also far more expensive than unlucky losses, because losses at least get investigated.
Make the defense public and make it cost something. When a good decision produces a bad outcome, someone senior stands up in a room with other humans in it and says, on the record - this was the right call, I would authorize it again with the same information, and here is why. Not in a private message. Not as a footnote. A costly, visible signal, because a cheap signal teaches nothing and everyone can tell the difference.
The last one is the hard one, and its difficulty is the entire diagnosis. Defending a loser feels like defending failure, which feels like tolerating failure, which no leader wants stapled to their name. So the defense is offered privately, where it is free, and consequently where it changes nothing.
But an organization is only as brave as its most recent public reaction to a bad outcome. Everyone watched. Everyone updated. The people who reasoned well and drew badly learned exactly what happens to them, and the people who reasoned badly and drew well learned that the reasoning was never the thing being measured.
Coda
There is no cosmic scorekeeper reconciling reasoning with results. Reality does not grade on process, and it does not know it is grading. The distribution simply pays out, indifferently, and hands us a single sample with no error bars and a great deal of narrative momentum.
Which means the symmetry has to be manufactured, deliberately, by people willing to say the unnatural thing out loud - that this was a good decision, that it went badly, and that both of those are true at once.
Do it once and it sounds like an excuse. Do it consistently, with receipts written in advance, and it becomes the only accountability that was ever worth the name.