You always knew it would break (you just never said so out loud)

The outage retro has a ritual, and if you've sat through more than three of them you already know its shape by heart. Something breaks. The room assembles. And somewhere in the first ten minutes, at least one person says a version of: "Honestly? I had a bad feeling about that module for weeks."

Nobody asks the obvious follow-up question, which is: did you say that to anyone, at any point, before it broke? Because if the answer is no — and it is almost always no — then what just happened in that meeting wasn't a memory. It was a manufacture.

The experiment that named the trick your brain just played

In 1975, Baruch Fischhoff ran an experiment that has never really stopped being relevant. He gave people descriptions of historical events — a battle, a diplomatic negotiation — along with several possible outcomes, and asked one group to estimate, in advance, how likely each outcome was. He then told a second group which outcome had actually happened, and asked that group to recall, or re-estimate, what they would have predicted beforehand. The people who knew the outcome consistently claimed they'd have rated the actual outcome as more likely than the people who genuinely didn't know it yet — and, more unsettlingly, people who'd made real predictions before learning the outcome misremembered their own prior estimates as closer to the truth than they actually were. Fischhoff's paper carried a title built to be quoted forever: "Hindsight is not equal to foresight." He called the underlying mechanism creeping determinism — once you know how something turned out, your mind quietly rebuilds the past as a story where that outcome was always the obvious one, and it does this without asking your permission or flagging that it happened.

The finding held up. A half-century of replication later, Neal Roese and Kathleen Vohs reviewed the accumulated evidence in a 2012 Perspectives on Psychological Science paper and described hindsight bias as operating on at least three separable levels — you misremember what you actually predicted, you come to see the outcome as having been inevitable, and you come to see it as having been foreseeable, by you specifically, even when it wasn't. It shows up in medical diagnosis, in courtroom judgments of negligence, in sports commentary, in stock-picking — and, this should not be a surprise by now, in engineering retros, every single time someone says "I always knew that service was fragile" about a service they mentioned to precisely nobody until it was on fire.

Roese and Vohs note something that makes the bias more, not less, unsettling: it is remarkably resistant to simply being told about. Warning people about hindsight bias in advance measurably reduces it in some settings and does almost nothing in others, particularly when the outcome carries real emotional weight — and an outage carries plenty. Knowing the mechanism exists does not reliably immunize a room against it in the moment, which is exactly why this can't be solved with a slide at the start of the retro reminding everyone to watch out for hindsight bias. The fix has to happen earlier than the retro, at the point where the claim is actually made, not afterward, at the point where everyone tries to remember whether they made it.

It's worth being specific about what "at the point the claim is made" actually rules out, because the natural workaround is tempting and doesn't work. Writing the risk down in a private notebook doesn't count, not because private notebooks are dishonest, but because a note nobody else could have checked at the time is functionally identical, from the retro's point of view, to a memory recovered after the fact — there's no way to distinguish an honestly early private note from one quietly backdated by a mind that's very good at backdating things without noticing it's doing so. The claim has to land somewhere shared, timestamped by something other than the person making it, before anyone in the room can treat it as evidence rather than as another retelling.

Why this isn't really about honesty

It's tempting to read this as a story about people lying to look smart after the fact, and sometimes that's exactly what it is. But the more important reading, and the one the research actually supports, is that the distortion is usually sincere. The person saying "I always knew" typically believes it completely, because their memory has already been quietly rewritten by the outcome they now know. That's what makes hindsight bias more dangerous than ordinary bragging: you can't argue someone out of a memory they experience as real, and a retro built on unrecorded recollection has no tool to tell a genuine early instinct from a retroactively installed one. Both feel, from the inside, exactly the same.

An unrecorded prediction isn't modest evidence of foresight. It's zero evidence, because your memory of having made it was assembled after you already knew the answer.

The only defense is a timestamp

This is the entire reason a claim only counts if it was said, quoted, attributed to a speaker, and logged before its horizon passed — never reconstructed afterward from anyone's confident recollection, however sincere. "I always knew" is worth nothing to a system that scores understanding. "I said, on the fourteenth, that the retry logic looked like the real cause" is worth everything, because it can be checked against what actually happened on the fourteenth, not against what everyone now agrees must have been obvious.

This cuts in a direction people don't expect the first time they hear it: it protects the person who was quietly, verifiably right just as much as it guards against the person falsely claiming credit. If you flagged the risky module in week two and nobody listened, a timestamped, quoted claim is the only thing standing between you and a retro where the room's collective hindsight bias erases your actual foresight along with everyone else's manufactured version of it. The room will remember it as "we all kind of knew." The quote will remember it as you, alone, on a specific date, before anyone else agreed with you.

Notice what this requires, and why it's stricter than it first sounds: not a paraphrase, not a summary written up afterward by whoever runs the retro, but the actual sentence, attributed to the actual speaker, timestamped before the horizon. A claim reconstructed after the fact from someone's confident memory of the meeting is exactly the artifact hindsight bias is best at producing — it will read as clean, prescient, and obviously correct, because it was assembled by the same mental machinery Fischhoff caught in 1975, working backward from an answer it already had. The verbatim requirement isn't bureaucratic caution. It's the only barrier standing between a real prediction and a very convincing forgery of one, and the forger, in this case, is nobody's enemy — it's everyone's own memory, doing exactly what memory does.

What to do with this on Monday

You don't need any product to start fighting this. The cheapest version is a running, dated list — a channel, a doc, anything with a timestamp nobody can quietly edit — where risk calls get written down the moment someone actually believes them, not the moment they're proven right. The rule that makes it work is brutal in its simplicity: if it isn't written down before the horizon, it doesn't count at the retro, no matter how certain the memory feels. Certain memories are exactly the ones hindsight bias is best at manufacturing.

There's a gentler way to think about why this rule is fair rather than punitive, and it's worth ending on. Nobody is being accused of dishonesty when a retro discounts an unrecorded "I had a feeling." The accusation, if there is one, is aimed at the format of the meeting, not the person in it — a standup that never asked anyone to write a risk down the moment they felt it is a standup that guaranteed, structurally, that Fischhoff's effect would eventually run wild in its own retros. Fix the format — a place to log the feeling the moment it happens, however small, however unofficial — and the hindsight problem doesn't need policing at all. The record does the policing on its own, gently, just by existing.

The uncomfortable mirror image is worth naming before closing: hindsight bias doesn't only inflate the credit people give themselves for foreseeing disasters. It works just as reliably in the other direction, quietly erasing the memory of confident predictions that turned out to be wrong. Nobody stands up at a retro and volunteers "actually, I was sure this would be fine, and I was completely wrong." The same creeping determinism that manufactures false foresight also launders away real overconfidence, and a record that only ever gets consulted to hunt for missed warnings is only doing half its job.