Fifty years of research keep saying the same thing: we decide badly, in predictable ways. Here are the seven works that matter, in plain words, each mapped to the exact Kapari feature that applies it.
Decision science is the study of how people and organizations choose, and of the regular mistakes they make while choosing. Since the 1970s its findings converge: human judgment is biased (it leans the same way, case after case) and noisy (two serious experts, the same file, two different conclusions). The fix is not demanding perfect decision-makers. It is equipping the process, by forcing disagreement into the open before the announcement.
Kapari is a test bench for decisions: you describe a decision, a panel of simulated voices grounded in real data reacts to it, and you see the range of plausible reactions before you commit. The theories below are not academic decoration: each one maps to a specific product feature, with its source.
And the difference with an expert opinion or an AI assistant answer comes down to one thing: the graded sheet is published. Across 86 real decisions replayed blind (43 French, 43 US), the engine found 77% of the objections actually documented at the time on the French exam and 81% on the US exam, worst sheet included, full file downloadable on the exam page.
A good decision can end badly, and a bad one can succeed: luck scrambles the grade. The authors draw the only workable rule: judge how the decision was made. Who argued the other side? Which uncertainties were actually explored? These questions almost never get asked in a meeting, because disagreement is socially expensive.
What Kapari codes from it: at the end of each run, the engine counts the substantive disagreements that appeared and the blind spots, meaning what no voice defended. Counts on panel data, not sentences written by an AI.
The full story: A good decision can end badly. So what do you grade?
"Noise" documents an uncomfortable fact: two serious judges, the same file, two conclusions. The remedy has been known since Meehl (1954) and Dawes (1979): a simple rule, applied consistently, holds its own against the expert case by case, precisely because it never gets tired and never changes mood.
What Kapari codes from it: the verdict is computed by published rules, and AI only gives the voices their texture. Same file, same verdict, replayable in front of a witness.
The full story: The verdict is computed, not generated by an AI.
The method fits in one sentence: "the project has failed, tell me why". By moving the team into a failure already taken for granted, you release the reservations politeness was holding back. Klein published it in the Harvard Business Review in 2007.
What Kapari codes from it: the premortem fires automatically when panel approval runs too wide. When everyone says yes is exactly when to look for what could break.
The full story: When everyone agrees, the engine raises a flag.
Two scientific adversaries ended up publishing together: intuition works in regular environments, where you practice with fast feedback. It derails on rare decisions, the kind an executive makes once in a career.
What Kapari codes from it: the rule applies to the tool itself. Kapari tells you when not to trust it: a topic outside its documented frames, missing context, a decision that is not one.
The full story: Our tool tells you when not to trust it.
The "outside view": instead of walking through your own plan, you compare your decision to the class of similar decisions already attempted, and look at how they turned out. Flyvbjerg turned it into the reference method for forecasting large projects.
What Kapari codes from it: a sourced yardstick that puts the decision back in its reference class, with the rule that goes with it: when no reliable benchmark exists, nothing is displayed.
The full story: Your case is less special than you think.
Twenty years of measuring expert judgment: those who know one big thing (the hedgehogs) get it wrong more often than those who know many small things (the foxes). What protects you is not one person's expertise, it is the diversity of angles.
What Kapari codes from it: the panel is composed to cover the camps, not to be big. Around thirty contrasted voices, each grounded in a documented profile, rather than three experts who agree with each other.
The full story: Why thirty voices beat three experts.
The founding paper of the field shows that a language model, conditioned on real profiles, can emulate response distributions close to those of human subgroups. The literature that followed measures the limits: instability, flattened opinions, gaps across subpopulations.
What Kapari codes from it: plausible voices to explore the structure of a decision, never an opinion percentage. Neither a poll nor a prediction: a range of reactions, with the limits published.
The full story: Are AI synthetic audiences accurate? And for the full market born of that paper: synthetic audiences, a map of the market.
Each detailed page in the series cites its source on screen and says what Kapari applies from the original work, and what it does not.
Kapari measures its audience with Google Analytics, to know how many people read these pages. Nothing is stored on your device until you accept.