↩ ExploRational · A Catalogue of Divergent Rationality
explorational · a divergent move
"Will this plan fail?" produces polite nods. "The plan has failed — it is a year later, and it was a disaster. Write the history of why." produces the list. Why does assuming the failure unlock reasons that merely fearing it cannot?
Gary Klein's premortem is a scheduled act of imagination. Before a plan launches, the team is told — not asked — that it has already died: it is a year from now, the project was a fiasco, and each person has two minutes to write its history. The move runs on prospective hindsight. Mitchell, Russo and Pennington found that imagining an outcome as certain and already-happened, rather than merely possible, increases the reasons people generate for it by roughly a third. A future that might fail is a fog of branching maybes; a past that did fail is a single corpse demanding a cause. Certainty converts prediction, which the mind defends against, into explanation, which the mind is greedy for.
The second mechanism is social, and it is the one meeting rooms actually starve for. While the plan is alive, naming its fatal flaw is an act of disloyalty — you are the one shooting at the thing everyone else is building. The premortem's premise inverts the etiquette: failure is stipulated, so criticism is no longer an attack on the plan but the assigned work of understanding a death. The assignment licenses dissent. Loyalty no longer requires silence; the junior engineer who privately knew the fatal flaw can finally say it, because she is not predicting doom — she is writing history. The instrument below runs both mechanisms as one disclosed model, and lets you watch the same room twice.
the signature instrument
A plan with 12 latent failure modes, each with an accessibility (how easily it comes to mind, blue) and a lethality (how badly it would hurt, red) — weakly anti-correlated by default, so the deadliest are often the least mentionable. The same k probes of imagination are run through two frames sharing the same random draws; the only differences are the optimism gate m and the dissent cost c. The displayed values are computed live from the seeded simulation.
The failure-mode archive
—
| failure mode | accessibility a | lethality l | prospective — "will it fail?" | premortem — "it failed. why?" |
|---|
"Will this plan fail?"
"It failed. Write the history of why."
the room · six advisors, one private doubt each · mechanism 2
—
—
certainty unlocks the archive
Ask "will it fail?" and you have asked the mind to bet against its own plan. The question runs through the optimism gate: each candidate reason must pass a motivated filter that admits bland worries easily and threatening ones reluctantly — in the model, a thought of lethality l survives with probability ml, so the gate leans hardest on exactly the thoughts that matter most. The probes still fire; the room hears "should be fine." The gated chips in the prospective column are the residue: reasons that came to mind and were swallowed.
Stipulate the failure and two things change at once. The filter loses its job — there is nothing left to protect, so m goes to 1 — and the search flips direction: instead of defending a prediction you are explaining a corpse, asking which causes a death like this one implies. Conditioning on failure makes a mode's lethality part of its relevance. That is why the effect in the literature is not subtle: across the studies behind prospective hindsight, certain-and-past framing lifted reasons generated by about a third — and the model is calibrated so its gate reproduces that lift (over 20,000 seeded runs at the defaults, +32.7% reasons, killer found 52% of runs prospectively against 65% premortem, on identical randomness).
what to try
Run both frames on the same seed and count what only the premortem finds. The frames share their random draws, so every green chip in the prospective column reappears in the premortem column — the premortem hears everything the polite frame hears. The difference is the red gated chips: probes that surfaced a mode and had it eaten. Reseed a few times; the gap in "lethality covered" is the price of the gate.
Drop the filter and the cost — press no filter anyway (m = 1, c = 0). The two columns become identical, chip for chip, and the room speaks 6/6 under either question. The premortem has no magic: it is only the removal of a filter and a fee. A team that already had neither gains nothing from the ritual — which is what the model should show.
Drag ρ positive and watch the move lose its edge. When the deadliest mode is also the most mentionable, the polite meeting nearly suffices — everyone was already thinking about the obvious cliff. The premortem earns its keep at ρ < 0, where the killer is quiet: accessible-but-mild worries soak up the prospective probes while the lethal mode waits in the archive.
the license to speak
The first mechanism lives inside one skull; the second lives between them. In the panel, each advisor holds one private doubt, and voicing it in a live-plan meeting costs standing c — dissent reads as disloyalty, so a doubt gets spoken only when its lethality visibly exceeds what the speaking costs: l ≥ c. Everything below that threshold stays in the hallway conversation after the meeting. Raise c to 0.95 — the polite meeting — and the room goes silent entirely: over the verification runs, the killer sat in somebody's head 36.5% of the time and was voiced in the meeting never.
The premortem's premise sets the same cost to zero. Once failure is stipulated, naming its cause is not an attack on the plan — the plan is already dead, and the person listing what killed it is doing the assignment. The same six people, holding the same six doubts, go from a couple of safe admissions to the full list, and nobody had to become brave. The move does not ask for courage; it re-prices it. The chips in the room panel show the flip directly — silent under the live plan, spoken over its body.
model
This is a behavioral page. The model is stylized-empirical — a disclosed mechanism built to reproduce a robust finding, like a disclosed anchoring model — not a derived law. Here is all of it. Each of 12 modes gets an accessibility a and a lethality l, drawn from a Gaussian copula with correlation ρ (default −0.40; the measured correlation over the verification runs was −0.39). Both frames search for causes of failure, so a mode's relevance is a·l in both. Each of k probes draws one shared uniform: the premortem maps it straight to a mode; the prospective frame maps it to the same mode, then passes it through the gate with probability ml, else voices nothing. The room speaks a held doubt prospectively iff l ≥ c, and always speaks it under the premortem premise.
One number is calibrated, and the calibration is the disclosure: m = 0.60 was chosen so the gate passes about 75% of probes at the defaults, making the premortem's expected lift in reasons ≈ +33% — the "roughly a third" from the prospective-hindsight studies is an anchor the model is tuned to, not a result it discovers. The rest — which modes surface, whether the killer is found, what the room says — falls out of the seeded draws you are watching. The real effect sizes vary by task and by team; treat the mechanism as the claim, and every specific percentage as illustration.
the move ↔ the model
| The model | The move |
|---|---|
| a failure mode | a reason the plan could die — present in the world whether or not anyone says it. |
| accessibility a | how easily it comes to mind: the worry you already rehearse versus the one nobody has words for yet. |
| the optimism gate ml | what fear of disloyalty and love of the plan suppress — leaning hardest on the deadliest admissions. |
| the premortem premise | certainty that removes the filter: with failure stipulated, there is nothing left for optimism to defend. |
| conditioning on failure (a·l) | asking which causes a corpse implies — lethality becomes relevance once the death is given. |
| the surfaced list | the history written in advance: the premortem's output, ready while the plan can still be changed. |
how this opening fails
risk · endless deferral
The catalogue's named risk for this move. An instrument this good at generating reasons to fail can become a machine for never acting: run it weekly and every plan is always freshly dead, every list a warrant for one more delay. Klein's version is a scheduled exercise with an end — two minutes of writing, one round of reading, then the plan is alive again and the list becomes mitigations with owners. The premortem is a diagnostic you perform on a launch, not a mood you keep the team in.
risk · the parameters are illustrative
m = 0.60 and the l ≥ c speaking rule are stylizations chosen to reproduce a documented lift of roughly a third; real filters and real meeting costs vary with stakes, culture, and who is in the room, and nobody has measured your team's m. And the model only ever samples from modes that exist in someone's archive: a premortem redistributes attention over traces people already carry. It cannot surface a failure mode nobody in the room has any trace of — for that you need different people, not a better question.
Declare the patient dead, and the room will finally tell you what was killing it.
can you use it?
RECOGNITION — Which is a premortem? A: reviewing what went wrong after the project failed. B: standing at launch and writing the history of a failure that has already happened — 'it's a year from now and this flopped; here's why.' C: listing risks by probability.
B. Prospective hindsight: the failure is assumed complete, which unlocks reasons optimism was suppressing. A is a real postmortem (too late); C is a risk register, not a story.
THE NEAREST NEIGHBOR — Consider-the-opposite also inverts the expected outcome. What separates the premortem from it?
Tense and completeness. Consider-the-opposite entertains the reverse of a belief; the premortem assumes a specific failure has already fully occurred and narrates its causes — the certainty is what licenses the honesty.
PRODUCTION — Pick a plan you're confident in. Write the first sentence of its premortem: it is one year later, the plan failed, and you are explaining the single most likely reason why.
A version: 'It's next spring and we killed the feature — turned out we'd built for the users who complained loudest, not the ones who pay.' Yours works if you wrote in completed past tense with a concrete cause, not a hedged 'might'.